CA2814625C - Method and apparatus for stabilizing production of pulp from wood chips - Google Patents

Abstract

Methods and apparatus for stabilizing production of pulp from wood chips, and more particularly for stabilizing dry-mass flow rate of wood chips to be fed from a discharging source to a chip refining line, through an upstream first chip processing stage, involve on-line measurements of wood chip properties, such as moisture content and bulk density, prior to the processing stage, to derive initial dry mass flow rate data related to the wood chips entering the processing stage. The stabilization approach further involves a predictive control model capable of deriving, from initial dry mass flow rate data, current speed setpoint data for a screw feeder provided a an entry of the refining line, according to a predetermined production rate. Thus, the speed of the screw feeder can be manipulated in accordance with the setpoint data to substantially stabilize the dry mass flow rate of the wood chips fed at the entry of the refining line. As a result, wood chip properties variability can be compensated at input of the refining line, so as to improve energy efficiency by achieving lower specific energy consumption in the refining stage.

Description

=
METHOD AND APPARATUS FOR STABILIZING PRODUCTION OF PULP
FROM WOOD CHIPS
Field of the invention The present invention relates to the field of pulp and paper, and more particularly to method and systems for stabilizing production of pulp from wood chips.
Background of the invention In the pulp and paper industry, a well known pulping process involves shredding and defibering chips of wood between the electrical-driven rotating discs (plates) of a device called refiner. The product is known as Refiner Mechanical Pulp (RMP). RMP typically retains more long fibres than stone ground-wood and yields stronger paper.
The basic RMP process has undergone extensive development in the past three decades. Most new installations now employ thermal and/or chemical pre-softening of the chips to modify both the energy requirement and the resultant fibre properties. For instance, if the chips are given a pressurized steam pre-treatment, the resultant product, called Thermo-Mechanical Pulp (TMP), is significantly stronger than RMP and contains very little screen reject materials.
Mechanical pulping processes have the advantage of converting up to 95% of the dry weight of the wood into pulp, but require prodigious amount of energy to accomplish this objective. The pulp results in a highly opaque paper with good printing properties, but the sheet is weak and discolored easily on exposure to light.
To achieve adequate sheet strength, it is often necessary to add long-fibred chemical pulp. Newsprint traditionally was made up of about 75% ground-wood and 25%
chemical pulp. Now, newsprint is almost made from 100% TMP.
Semi-chemical pulping combines chemical and mechanical methods.
Essentially, the wood chips are partially softened or digested with chemicals.
The remainder of the pulping action is then performed mechanically, most often by disc refiners. Semi-chemical methods encompass the intermediate range of pulp yields between mechanical and chemical pulping i.e. 55% to 85% on dry wood. The pulps have a number of end-uses and some unique properties. As the prime example, pulps of about 75% yield exhibit exceptional stiffness, making them ideally suitable for the center fluted layer in corrugated container board.
Depending on raw materials and pulping process, pulps with a wide range of properties may be obtained. A large number of testing methods are in common use

2 to characterize pulps with respect to quality, processability, and suitability for various end uses. The more fundamental measurements provide the means to predict behavior, but are often too complicated to be applied in the mill laboratory.
Functional tests are designed to measure specific properties.
In terms of volume, newsprint is the most important member in the family of non-coated wood-containing printing papers, which also includes catalogue and directory papers. In the broadest sense, newsprint can be defined as any paper capable of being run through a modem high-speed printing press and producing an acceptable newspaper sheet at reasonable cost. The functional requirements of newsprint are pressroom runability, printability, good general appearance (i.e.
brightness, whiteness, cleanliness, opacity) and low price.
Newsprint furnish is made up mainly of mechanical pulp and/or recycled deinked newspaper stock, commonly in admixture with a small percentage of lightly-refined chemical pulp. The only general additive is violet-blue dye or so-called Bleaching Agent to offset the natural yellow hue of most mechanical pulps. In some cases, a small amount of mineral filler may also be added.
The mechanical pulp contributes such valuable printability-related properties to the newsprint as absorbency, bulk, compressibility, opacity, formation (i.e.
grammage micro-uniformity), etc. Unfortunately, the strength properties of the mechanical pulp may be insufficient to produce a sheet that runs well on the presses.
For instance, if stone groundwood is the only mechanical pulp used, the paper is usually reinforced with up to 30% chemical pulp, either semi-bleached Kraft or unbleached sulphite. While giving the newsprint greater strength, the chemical pulp adversely affects printing characteristics.
IMP is the most energy intensive pulping processes with the main advantage of converting up to 95% of wood dry weight into pulp. Almost 88% of the required energy is consumed in refiners for wood chips defiberization. The share of energy, water and vehicle fuel in final production cost accounts for about 21% of the total cost.
Theoretically, a small amount of energy is needed to create new fibre surfaces. In a typical process, more than 95% of electrical energy in refiners is transformed into thermal energy in the form of hot saturated steam. The majority of energy efficiency projects in IMP focus therefore on improving refining energy consumption. A review of the efforts already carried out in the last 20 years is presented by: Jackson and al. in "Mechanical Pulp Mills; Energy Cost Reduction in the Pulp and Paper Industry" Browne, T.C. tech. ed., Paprican (1999); Nilsson et al.

3 "Energy Efficiency and the Pulp and Paper Industry" American Council for an Energy-Efficient Economy, Berkeley, CA (1995); Corson et al. "Mechanics of Making Better Pulp with Less Energy" Pulp Pap. Int. 39(5):61-65 (1997); and Browne et al.
"Energy cost reduction in the pulp and paper industry: an overview" Pulp Paper Can.
102(2):T34-T38 (2001). The TMP mills are large consumers of electrical energy.
Disc refiners, typically powered by large 10-30 MW electric motors, are used to convert wood chips to high quality papermaking fibres. According to analysis results of M.
Jackson et al., the energy consumption for a 500 BDMT/D (Bone Dry Metric Ton per Day) single-line TMP mill at 2400 kWh/130MT, which is typical for a TMP mill using black spruce chips for newsprint production, was estimated at 2160 KWIVADt (K Watt-hourper Air Dry ton) which corresponds to 90% of the whole mill energy consumption. Since the TMP process is used in 80% of the newsprint production worldwide, energy consumption is a major issue in that industry.
Recent studies demonstrate the direct impact of wood chips properties including wood species, degree of freshness, moisture content, dimension profile, basic density, bulk density, fibre length and seasonal variations on the required refining energy and the quality of the pulp, as reported by: Brill "Effects of wood and chip quality on TMP properties" Proceedings International Mechanical Pulping Conference, pp. 153-161 161, Stockholm, Sweden, 1985; Rudie et al "The influence of wood and fiber properties on mechanical pulping. TAPP! J., 77, pp. 86-90, 1994;
Varhimo et al. "Mechanical Pulping" Vol. 5, Papermaking Science and Techonology Series., chapter 5: Raw Materials, Finnish Paper Engineers' Association, TAPPI, pp.
66-104,1999; Ding et al. "Effects of wood chip characteristics on refining energy consumption" International Mechanical Pulping Conference (IMPC), Minneapolis, USA, September 2007; and Smock "Handbook for Pulp & Paper Technologies", Joint Textbook Committee of the Paper Industry, 54 (1982). Any variation of the chips properties is propagated as a result through the pulping process that causes process variability. More refining energy is consequently required to compensate the effects of the process variability especially on the drainability or freeness CSF
(Canadian Standard Freeness) as the main pulp quality index in TMP. The secondary side-effect is more severe constraints on process parameters that result in less efficient operation.
The relations between the refining process and pulp quality have been exhaustively discussed by Miles in "Refining Intensity and Pulp Quality in High-Consistency Refining", Paperi ja Puu ¨ Paper and Timber, 72(5): 508-514, (1990), by Stationwala et al. in "Effect of Feed Rate on Refining'. Journal of Pulp and Paper

4 Science: vol 20 no 8 (1994) and by Wood in "Chip Quality Effects in Mechanical Pulping ¨ A Selected Review" 1996 Pulping Conference pp. 491-495. Furthermore, the relations between refining process and chip properties have also been exhaustively discussed by Jensen et al in "Effect of Chip Quality on Pulp Quality and Energy Consumption in RMP Manufacture", hit symp. on fundamental concepts of refining, Appleton Wis., sept. (1980), by Breck et at. in "Therrnomechanical Pulping ¨
a Preliminary Optimization", Transactions, Section technique, ACPPP, 1-3, pp (1975) and by Eriksen et at. in "Consequences of Chip quality for Process and Pulp Quality in TMP Production", International Conference, Mechanical Pulping, Oslo, June (1981).
According to a known control strategy, a feedback controller is used on the chip transfer screw feeder to control primary motor load, the dilution flow rate for the primary refiner being coupled with the screw feeding to operate on a constant ratio mode. Alternatively, the feedback controller can be used to control the motor load by acting upon the dilution flow rate on the basis of a pulp consistency measurement at the blow line of the primary refiner. In both cases, the variation of chip quality acts as an external disturbance affecting the motor load.
Most of existing control strategies in IMP function on the basis of the mean value of the chips properties entering the refiners over a long period of time and ignore the variations through time in operating the process, since they generally do not use on-line real-time measurements.
A system for measuring optical reflection characteristics of chips such as brightness, along with other important chip properties, such as moisture content, which is commercially known as the Chip Management System (CMS), is described in U.S. Patent no. 6,175,092 B1 and U.S. Patent no. 7,292,949 B2 both issued to the present assignee. Some pulp mills have used such system to manage their chip piles according to chip quality, as discussed by Ding et al. in 'Economizing the Bleaching Agent Consumption by Controlling Wood Chip Brightness" Control System 2002 Proceedings, June 3-5, 2002, Stockholm, Sweden, pp. 205-209. Chip quality assessment can be defined as the synthesis of measurements made of chip physical characteristics, as explained by Ding et al. in 'Effects of Some Wood Chip Properties on Pulp Qualities" 89th Pulp and Paper Annual Conference Proceedings, January 29, 2003, p. 35. Ultimately, this definition depends on the importance of each chip characteristic for a given process. Continuous variations in wood basic density and moisture content occurring in chip flow tend to cause variations in refining consistency, which in turn affect pulp uniformity and energy consumption as reported in U.S. Patent No. 7,292,949 82 and by Ding at aL in " Wood Chip Physical Quality Definition and Measurement", 2003 International Mechanical Pulping Conference, Quebec City, Canada, June 2-5, 2003, pp. 367-373 in view of TyvaInen "The Influence of Wood Properties on the Quality of TMP Made from Norway Spruce (Picea abies) - Wood from Old-growth Forests, First-thinnings, and Small Chips*
1995 International Mechanical Pulping Conference, pp. 23-34, 1995.
Many studies have shown that wood species and dry-based density are the dominant factors in pulping performance and pulp quality. The spruce family is the most favorable species for TMP as mentioned by Varhimo, A. at al, in "Raw Materials' in Sundbolm, J. "Mechanical Pulping' Chapter 5, Fapet OY, 66-104 (1999). Although chip aging can be observed from chip brightness, it is only useful for substantially unvaiied wood species. When an unknown proportion of wood species is present, more information is needed to provide reliable chip quality assessment.
Basic density is one of the most studied wood properties, and it varies substantially between and within various wood species. Basic density is not, however, an independent property but is determined by several characteristics of wood. As also mentioned by Varhimo et al, variations in wood basic density result in pulp quality variations. There is a good correlation between basic density and dry bulk density, the chip flow usually is metered by volume, and dry bulk density variations will cause fluctuations in the production rate, as reported by Dundar, E. at al, in "Decreasing Specific Energy of Thermomechanical Pulps from Reduction of Raw Materials Variability", September 2009, TAPP! Journal pp.23-29.
In prior published U.S. Patent application no. 2007/0158040 also naming the present assignee, there is disclosed a method and system making use of chip properties measurements to estimate the optimal dosage of bleaching agent for the purpose of control thereof in a pulp production process, by modeling the relationship between wood chips quality and pulp brightness. In particular, the model is used to evaluate the minimum charge of peroxide required to reach certain level of pulp brightness according to possible chips properties fluctuations, in order to minimize the cost and environmental impact of the bleaching operation.
An on-line measurement system such as described in U.S. Patent no.
7,292,949B2 and Ding et al. cited above can produce data that is useful for identifying the proportion of pure wood species making up a mixture of wood chips, on the basis of optical reflection and moisture measurements made on wood chips.
US 7,292,94982 teaches the use of an estimation model based on a feed-forward neural network that is built from optical reflection-based measurements, namely and dark chip content (D), along with moisture measurement as input variables, in which chip freshness (ageing) and species are controlled, and the selection of the input variables for the FFNN has been performed using known Principal Component Analysis (RCA) technique from the trials results. However, it has been observed that such approach provides an estimation of the proportion of each species within a range of only about 10%, which is generally insufficient to allow an efficient control over species variation in wood chips fed to the pulping process. Although US 7,292,949132 teaches that chip quality on-line measurement is very useful for stabilizing chip input, and that feedback information can control chip-feeding screws so as to take suitable proportions of chips from different piles or silos, such approach is not efficient to minimize specific energy. The system described in US 7,292,94982 for measuring chip moisture content and wood species does not involve any chip density measurements.
In prior published U.S. Patent application no. 2006/0278353 also naming the present assignee, there is disclosed the use of a measurement system for estimating and controlling relative proportion of wood chips originating from a plurality of sources characterized by various wood species, in a mass of wood chips to be fed to a process for producing pulp, wherein light reflection-related and density-related properties are used as input in a model characterizing a relation between such wood chip properties and species information. This principle allows efficient monitoring of the variation in wood species composition characterizing the wood chips to be processed, for the purpose of stabilizing chip feeding control and optimizing process parameters adjustment. When installed in the chip feeding process, the measurement system generates on-line chip characteristics information that can be used to control the mixture of chips from the different piles in order to stabilize the dry mass of wood chips entering the digester stage of a Kraft process. Such approach is also discussed by Ding et al. in 'Reduction of TMP refining energy by stabilizing wood chip basic density" PTS Paper symposium, Kleine Kongresshalle, Munich/Germany, September 7-10, 2010.
As opposed to chemical processes involving digesters, pulp production processes such as TMP, CTMP or MP generally use, upstream the refining process, processing stages of retention, atmospheric presteaming, washingidewatering, preheating and/or impregnation (CTMP), etc. Similarly, production processes of wood-based products such as MDF or HDF generally use, upstream the refining process, processing stages of retention and preheating/steaming. Due to these upstream processing stages, the known, direct chip feeding control approaches used in chemical processes involving digesters cannot be used to efficiently stabilize the refining process. In the system disclosed in prior published U.S. Patent application no. 2011/0264258 also naming the present assignee, where a chip processing stage upstream of a chip refining process is involved, dry-based density data, moisture content and one or more light reflection-related properties are used by a reference model for selectively modifying the discharge rate set points of one or more of the wood chip sources, to provide substantial stabilization of the dry-based density of the wood chips around a predetermined target. Although an improvement over prior process stabilization approaches, there is still a need for further enhancement of refining process efficiency.
Presently, variations in specific energy consumption (SEC), i.e. applied energy per unit of weight of wood chips on an oven-dry basis during refining, to obtain a desired pulp quality can be relatively high. Usually there is a range of desired quality values, such as provided by Canadian Standard Freeness (CSF) for example, with which the produced pulp must comply to satisfy customers' demand. In this range, the obtained CSF can sometimes be near the upper limit or the lower limit.
When the value is near the lower limit of the desired range, this means that more energy is needed to reach the desired quality. When the value is near the upper limit, a minimal consumption of energy for an acceptable quality pulp is reached. For cost reduction and resource protection purposes, it is desirable that energy spent to produce a pulp of a desired quality is managed efficiently. Refiners are also involved in the manufacturing of fibreboards made from various lignocellulosic granular matters including wood chips and mill waste matters such as wood shavings, sawdust or processed wood flakes (e.g. OSB flakes). While the respective post-refining steps of fibreboard manufacturing and pulp and paper processes are distinct, their refining modes of operation are similar, and cost reduction as well as resource protection are important issues for both processes, so that it is still desirable that energy spent to produce a pulp of a desired quality is managed efficiently.
In prior published U.S. Patent application no. 2010/0121473 also naming the present assignee, there is disclosed a method and a system for optimizing the operation of a refining process, making use of a predictive model including a simulation model for the refining process which is based on relations involving a plurality of lignocellulosic matter (e.g. wood chips, wood shavings, sawdust or processed wood flakes) properties to be fed to the process, input operating parameters (e.g. chip matter transfer screw speed, dilution flow rate, hydraulic pressure, plate gaps, retention time delays), controlled output parameter (e.g.

primary motor load, pulp freeness) and specific energy consumption (SEC) as uncontrolled output parameter, to generate a predicted value of SEC. There is provided a simulation model adaptor fed with data representing measured values of the matter properties and data representing measured values of the controlled and uncontrolled output parameters, to adapt the relations of the simulation model accordingly. An optimizer is provided for generating an optimal value of control target according to a predetermined condition (minimization) on predicted SEC and to predetermined process constraints. Such approach is also discussed by Dundar et al.
in 'Effects of wood chip characteristics on refining energy consumption' 2007 International Mechanical Pulping Conference, Minneapolis, MN, 6-9 may 2007.
The potential of SEC reduction through low variability (LV) analysis based historical data of a TMP process is discussed by Shahriari et al. in "Assessing Potential Specific Energy Reduction in a TMP Process" Proceedings of EXFOR conference, Montreal, QC, Canada, Feb. 2-3 2010. Shahriari et al. in "Investigating potential energy savings from raw materials variability attenuation: A case study in pulp and paper industry", 2011 24th Canadian Conference on Electrical and Computer Engineering (CCECE), Niagara Falls, (ONT), Canada 8-11 May 2011, scrutinized the impact of wood chips properties (WCP) variation on process variability in a TMP process, on the basis of a starting hypothesis stating that by reducing WCP variation, process variability is mitigated, that results in energy efficiency and product quality enhancements.

Process data during two months of continuous operation of the process were analyzed and low variability periods were identified. The periods were then compared to the average of the process to characterize the correlation between WCP
variation, energy efficiency, and product quality indexes. As a result of complementary works, Shahriari et al. in "Process variability and inherent efficiency enhancement in industrial processes: Two case studies in pulp and paper industry" 2011 IEEE
International Conference on Control Applications, Quebec, QC, Canada, 28-30 Sept 2011, further highlighted the correlation between process variability attenuation and efficiency enhancement in industrial processes such as TMP process. On the basis a multi-variable statistical approach requiring no process model, the hypothesis that process variability attenuation brings about efficiency enhancements in both energy and productivity aspects, was further confirmed.
Several research projects have already been carried out to develop physical models of TMP, mainly to optimal control of the process, such a reported in the following publications: Qian at al. "Mechanistic model for predicting pulp properties from refiner operating conditions" TAPPI Journal, 78:215- 222, 1995; Lama "Controllability Limitations of TMP-Refining Process" PhD thesis, Ecole Polytechnique de Montreal, 2006; Harinath et at. "Control and optimization strategies for therrno-mechanical pulping processes: Nonlinear model predictive control"
Journal of Process Control, 21(4):519- 528, 2011 ;Tervaskanto et at. "Refiner quality control in a CTMP plant" Control and Automation, IEEE International Conference on ICAA 2009, pp 1266-1271, dec. 2009; Biegler et at. "Advanced step nonlinear model predictive control for two-stage therm mechanical pulping processes"
Proceedings of the 18th IFAC World Congress, 2011; Strand " Simulation, control and on-line optimization on mechanical pulping systems" PhD thesis, Idaho University, 2008;
Olen Modelling and Dynamic Simulation of a CTMP
Plant" PhD thesis, The University of British Columbia, 1997; Kidd"
Implementation of a TMP advanced quality control system at a newsprint manufacturing plant*
Technical report, Augusta Newsprint Company, 2006; Du " Muftivariable Predictive Control of a TMP Plant" PhD thesis, The University of British Columbia, 1998;
Schwartz et at. " A method of modeling, predicting and controlling TMP pulp properties", Proceedings of the 1996 IEEE International Conference on Control Applications, pp. 846-851, sep 1996; Lappalainen " Validation of plant dynamic model by online and laboratory measurements a tool to predict online COD loads out of production of mechanical printing papers" PhD thesis, Lappeenranta University of Technology, 2008; and Di Ruscio "Topics in Model Based Control with Application to the Thermo Mechanical Pulping Process' PhD thesis, The Norwegian Institute of Technology, 1993. 42. However, the developed models and control strategies mostly remain in academic while few are implemented in industry.
Parallel to physical modeling, numerous numerical/statistical/empirical approaches have also been developed to model TMP. The main distinction between physical and numerical models is that the latter are aimed to be used in advanced control strategies, as discussed in the following publications: Du et al.
"Nonlinear control of a wood chip refiner Proceedings of the 4th IEEE Conference on Control Applications, pp. 1065-1066, sep 1995; Di Ruscio "Model predictive control and identification: a linear state space model approach' Proceedings of the 36th IEEE
Conference on Decision and Control, volume 4, pp. 3202-3209, dec 1997; Elsinga "TMP optimization using multivariate analysis., IEEE Transactions on Industry Applications, 39(3):893-898, may-june 2003; Hietanen et al. "The process control using SPC and fuzzy modeling techniques" IFAC 15th triennial world congress, Barcelona, Spain, 2002; and Angelou et al. "Artificial intelligence tools for the on-line prediction of quality properties in pulp and paper processes" Paper Summit, Atlanta, GA, 2002. The line between physical and empirical models is not distinct and a grey region exists in between. An example is the model presented by Lama cited above, that is a combination of physical models for mechanical elements and a neural-net based empirical model for pulp quality.
Summary of the invention The general object of the proposed invention is to compensate wood chip properties variability at the input of a refining line, so as to improve energy efficiency by achieving lower specific energy consumption in the refining stage.
According to the above general object, from a broad aspect, there is provided a method for stabilizing dry-mass flow rate of wood chips to be fed from at least one discharging source to a chip refining line provided with a screw feeder at an entry thereof, through a first chip processing stage upstream said refining line.
The method comprises the steps of: 0 estimating on-line a set of wood chip properties characterizing said wood chips prior to be fed to said first chip processing stage, to generate corresponding wood chip properties data, said set including moisture content and bulk density to derive from said wood chip properties initial dry mass flow rate data; ii) feeding said initial dry mass flow rate data at an input of a predictive control model capable of deriving therefrom current speed setpoint data for said screw feeder according to a predetermined production rate of said refining line; and iii) manipulating the speed of the screw feeder in accordance with the setpoint data to substantially stabilize the dry mass flow rate of the wood chips fed at the entry of the refining line. In an embodiment, the predictive control model includes a first model representing said processing stage for estimating a lag time between said estimating step i) and the wood chips feeding at the entry of the refining line to generate predicted dry mass flow rate data; and a second model representing said screw feeder and receiving said predicted dry mass flow rate data and said predetermined production rate to derive therefrom said current speed setpoint data.
According to the same general object, from another broad aspect, there is provided a method for stabilizing production of pulp from wood chips to be fed from at least one discharging source to a chip refining line provided with a screw feeder at an entry thereof, through a first chip processing stage upstream said refining line. The method comprises the steps of: i) estimating on-line a set of wood chip properties characterizing said wood chips prior to be fed to said first chip processing stage, to generate corresponding wood chip properties data, said set including moisture content and bulk density to derive from said wood chip properties initial dry mass flow rate data, said set further including at least one reflectance-related property; ii) feeding said initial dry mass flow rate data at an input of a first predictive control model capable of deriving therefrom current speed setpoint data for said screw feeder according to a predetermined production rate of said refining line;
iii) controlling the speed of the screw feeder in accordance with the setpoint data to substantially stabilize the dry mass flow rate of the wood chips fed at the entry of the refining line; iv) feeding the wood chip properties data at an input of a further predictive control model capable of deriving therefrom further current setpoint data for an associated control parameter for controlling a quality parameter characterizing the pulp to produce from said wood chips; and v) manipulating the control parameter in accordance with the further setpoint data to substantially stabilize the pulp quality parameter. In an embodiment, the further predictive control model uses said first processing stage model for estimating the lag time between said estimating step i) and the wood chips feeding at the entry of the refining line to generate predicted wood chip properties from which said further current setpoint data are derived. In another embodiment, the associated control parameter is bleaching agent flow rate, and the pulp quality parameter is pulp brightness.
According to the same general object, from another broad aspect, there is provided an apparatus for stabilizing dry-mass flow rate of wood chips to be fed from at least one discharging source to a chip refining line provided with a screw feeder at an entry thereof, through a first chip processing stage upstream said refining line.
The apparatus comprises means for estimating on-line a set of wood chip properties characterizing said wood chips prior to be fed to said first chip processing stage, to generate corresponding wood chip properties data, said set including moisture content and bulk density to derive from said wood chip properties initial dry mass flow rate data; a control unit provided with a predictive control model receiving at an input thereof said initial dry mass flow rate data, for deriving therefrom current speed setpoint data for said screw feeder according to a predetermined production rate of said refining line; and means for manipulating the speed of the screw feeder in accordance with the setpoint data to substantially stabilize the dry mass flow rate of the wood chips fed at the entry of the refining line.
Still according to the same general object, from another broad aspect, there is provided an apparatus for stabilizing production of pulp from wood chips to be fed from at least one discharging source to a chip refining line provided with a screw feeder at an entry thereof, through a first chip processing stage upstream said refining line. The apparatus comprises means estimating on-line a set of wood chip properties characterizing said wood chips prior to be fed to said first chip processing stage, to generate corresponding wood chip properties data, said set including moisture content and bulk density to derive from said wood chip properties initial dry mass flow rate data, said set further including at least one reflectance-related property; a control unit provided with a predictive control model receiving at an input thereof said initial dry mass flow rate data, for deriving therefrom current speed setpoint data for said screw feeder according to a predetermined production rate of said refining line; means fbr manipulating the speed of the screw feeder in accordance with the setpoint data to substantially stabilize the dry mass flow rate of the wood chips fed at the entry of the refining line; a further control unit provided with a predictive control model receiving at an input thereof said wood chip properties data, for deriving therefrom current setpoint data for an associated control parameter for controlling a quality parameter characterizing the pulp to produce from said wood chips; and means for manipulating the control parameter in accordance with the further setpoint data to substantially stabilize the pulp quality parameter.
In an embodiment, the associated control parameter is bleaching agent flow rate and the pulp quality parameter is pulp brightness.
Brief description of the drawings Some embodiments of methods and apparatus according to the present invention will be described below in view of the appended drawings, in which:
Fig.1 is a schematic representation a typical TMP process to which the present invention may be applied;
Fig. 1A is a schematic block diagram of a first control unit using a closed-loop feedback system to stabilize dry bulk density at wood chip discharging stage;
Fig. 1B is a schematic block diagram of a second control unit making use of a predictive control model to generate current speed setpoint data for the screw feeder of the refining line for purposes of dry mass flow rate stabilization;
Fig. 1C is a schematic block diagram of a third control unit illustrated in Fig.
1C, making use of of a furthz r predictive control model that can be used to stabilize pulp quality;
Fig. 2 is a detailed block diagram representing an control strategy integrating dry mass flow rate and pulp quality stabilization;
Fig. 3 is a simplified representation of the chip processing stage and refining line that can be used for modeling purposes;
Fig. 4 is a theoretical representation of two typical tanks in series as typically involved in the processing stage for modeling purposes;
Fig. 5 are graphs presenting the variation of dry bulk density and the impact on the motor loads of the primary refiners as measured on two refining lines;

Fig. 6 is a graph presenting the results of a batching and averaging procedure for chips moisture content data;
Fig. 7 includes three graphs presenting the results of error modeling according to an example involving three proposed modes;
Fig. 8 is a graph showing the result of a simulation of dry bulk density prediction in primary refiners;
Fig. 9 is a graph comparing predicted chip moisture contents at constant and variable production rates;
Fig. 10 is a graph presenting the results of error modeling according to an example involving a mode wherein a mixing/segregation phenomenon is taken into account;
Fig. 11 is a graph showing the results of the model prediction in three modes related to the example of Fig. 10;
Fig. 12 includes graphs showing validation curves for prediction model parameters related to dry bulk density and refiners motor loads;
Fig. 13 is a graph showing a distribution of lag time prediction for an example of identification and validation data-sets;
Fig. 14 is a schematic block diagram of input/output flows in a latency chest Fig. 15 and 16 are graphs showing the variability of analyzed variables with respect to a stability threshold in identification data-set;
Fig. 17 is an example of multi-layer perception neural network that can used for modeling;
Fig. 18 is an example of radial-basis function neural network that can used for modeling;
Fig. 19 is an example of recurrent neural network that can used for modeling;
Fig. 20 to 24 are graphs showing the performance results of an experience for modeling with a multi-layer perception neural network for freeness as performed on a refining line respectively for time domain, mean, standard deviation, best and regression;
Fig. 25 to 29 are graphs showing the performance results of an experience for modeling with a radial-basis function neural network for freeness as performed on a refining line respectively for time domain, mean, standard deviation, best and regression;
Fig 30 is an example of decomposition of a pulp model;
Fig. 31 includes graphs showing distribution of error for predicted and corrected pulp quality indexes; and Fig. 32 includes graphs showing some degradation of predicted results when using a smaller horizon.
Detailed description of embodiments of the invention In the context of the present invention, a framework for refining energy optimization is proposed to enhance energy efficiency, productivity, and product quality of processes in pulp and paper industry. One of the approaches proposed in this framework is based on the assumption that efficiency, productivity and quality enhancements can be achieved through process stabilization by mitigating the variability of raw materials properties. This requires on-line real-time measuring of wood chips properties that can be carried out by employing existing chip measurement systems such as the Chips Management System (CMS) described in U.S. Patents nos. 6,175,092B1 and 7,292,949B2 naming the present assignee, together with Chips Weighing System (CWS) described in published U.S. Patent applications nos. 2006/0278353 and 2010/0121473 also naming the present assignee, which systems are basically used to measure basic density, bulk density, moisture content and brightness of the wood chips, and are available from the present assignee. The system has already been installed in several mills to manage the chip piles before the process or to control the bleaching agent in the process.
Referring now to Fig. 1, there is schematically shown an example of a typical TMP
process generally designated at 10, from the chips piles to the end of the final pulping stage before paper machines (not shown). Wood chips in the first place are classified and stored in different piles contained in silos 11, 12, 13 with respect to wood species, chips density and the degree of freshness. An optimal mixture of chips can be produced using an appropriate proportion of chips discharged to a main conveyor 22 from piles contained in silos 11, 12, 13 by operation of respective dosing screws 14,15,16. Although this proportion may be fixed at a predetermined value, it can be advantageously controlled using on-line real-time chip properties measurements using CMS/CWS (hereinafter referred to as CMS) as designated at 54, through first a first cont1 unit designated at 56 operating in feedback control loop to maintain the dry bulk density (DBD) of the mixture around a given setpoint.
However, the obtained level of process stabilization after initial chips discharge may still be improved later on before the chips enter the refining line, as will be explained later in detail.
Depending on the required pulp quality, an appropriate recipe is prepared and sent through an elevator 20 to a first processing stage 17, making use of a pre-heater bin 18 where the chips are heated for a predetermined period of time, typically of 30-90 minutes. They are then washed and screened using a chip washer 19 to make sure most of foreign materials are removed, and the extra water is drained with a screw drainer 21. The resulting wet materials are supplied to a retention bin 23 followed by a steam vessel 24 where they are softened with pressurized steam to allow a more easy defiberization with less fibre fracture during the refining process.
The softened chips fed with a continuously flow to a three-stage refining line through a plug screw feeder 27. In a IMP, refiners are rotating units propelled by electrical motors and provided with hard steel alloy plates, wherein electrical power is converted into mechanical and thermal energy by the action of the plates upon the chips, which are progressively broken down to smaller bundles and then into single fibres. To prevent fiber damages and bum due to friction and heat, dilution water is supplied between the plates and is mainly converted into steam. The water tends to reduce the consistency of the pulp as well. The pulp exiting from the first stage, primary refiner 31 passes through a first blow cyclone 28 to separate the steam from the fibers. The partially refined pulp is driven to the second stage refiner 32 and second blow cyclone 29, and then to a third stage refiner 33 and third blow cyclone 30, whereby the specific surface of the filers is further developed. After the third stage refining, fibers are twisted, kinked, and curled. To relax the fibers, the pulp is submerged in hot water inside a latency chest 35, where bleaching agent can also be added through a pump 62, as will be explained below in more detail. The refined pulp normally contains undesirable components such as grits and coarse fragments that can be removed by means of series-connected centrifugal screens 37, 38, sending the undesirable components to a reject chest 26 entry of a reject line 40 where rejects are processed by means of series-connected reject refiners 42, 43 followed by a reject latency chest 45 and a reject screen 47. Accepted pulp from screen and reject line 40 is sent at final pulping stage output 50 to the paper machines (not shown), while the rejects are removed from the process at reject line output 52 and are valorized in an alternative way. Hot steam can be partially recovered at an outlet 48 and recycled in other mill sectors such as paper machines to dry produced paper.
In a TMP, three principal quality indexes are freeness (CSF), fiber length (FL) and brightness (Bri). Freeness is the inverse of the resistance of a fiber mat to the flow of water that changes with respect to the amount of accumulating fines fraction.
It has a quasi-linear inverse relationship with refining energy. As for fiber length, it is largely affected by wood species, refiner's plate gap and dilution water flow rate.
Brightness is directly linked to the wood chips species and freshness. All three indexes are measured after the latency chest 35 and before the paper machines by a Pulp Quality Monitor (PQM) system designated at 57 in Fig. 1.
As mentioned above, WCP fed to the process evolve through time mainly due to wood species, wood age, growing environment and seasonal changes that affects moisture content, bulk/basic density and brightness of the chips. If this evolution is ignored, it provokes variation in the pulping process. The most important factors causing process variability are wood chips moisture content, bulk density and brightness. As to chips moistme content and bulk density, the main production constraint in a IMP process, apart from the refiners' motor load stability, is to keep the freeness in an acceptable range. For a given CSF value, a specific amount of refining energy (SRE) is required. Given the fact that the plug screw feeder 27 is volumetric, if the moisture content or bulk density of the chips varies, the dry mass flow rate fed to the refiner also changes. This results in variation in the applied SRE if the motor loads are kept constant which provokes CSF variation. In this perturbed environment, more SRE should be applied to stabilize CSF, to respect its upper limit and to keep it in the required range. More the process is variable, more extra SRE is required to compensate the variation. Regarding wood chips brightness, the wood species and chips freshness have direct impact on the brightness of the produced pulp. To respect the minimum required brightness, the bleaching agent is added to the pulp in the latency chest Likewise moisture content and bulk density variation, more the brightness of the chips varies, more extra bleaching agent is required in the latency chest to guarantee the minimum brightness of the pulp, which results in additional final product costs.
If chips classification is performed with respect to wood species, chips bulk/basic density and the degree of freshness represented by moisture content, an optimal chips recipe is produced using a portion of each pile through dosing screws.
In the lack of measuring unit, this portion is fixed considering the mean value of the chips bulk density and moisture content in each pile over a long period of time.
However, this solution is not robust against WCP variations through time. A
more detailed schematic block diagram of the first control unit 56 is illustrated in Fig. 1A, wherein the closed-loop feedback system is proposed to stabilize the dry bulk density (DI3D) of the produced recipe defined as:
DBD(t) = Hulk Densily(t) x (1 Moisture ConlenuO) (1) This stabilizes the applied SRE and consequently, the CSF as the main pulp quality index. The proposed scheme requires two models: a first model 47 representing the mass flow rate of each dosing screws 14, 15, 16 with respect to their rotational speeds, and a second model 51 representing the delay time of the conveyor from the output of the piles contained in silos 11,12,13 to the measuring point. The control unit 56 calculates the appropriate rotational speed of each dosing screw to reject the perturbations in considering the demanded production rate as reference. The problem to be solved is multi-variable control of a time delay system. The proposed strategy takes into account WCP high frequency variation compared to delay time of the conveyor and non-uniform mixture of wood chips on the conveyor that may result in imprecise WCP measurements that would otherwise cause instability of the closed-loop feedback control. In the present example, the focus is to stabilize chips dry bulk density (DBD reference).
However, in most of TMP mills, chips classification is not performed and the above control strategy may not be implemented alone. In a perturbed environment that DBD(t) vary through time, to stabilize the applied SRE to avoid CSF
variation, either refiners' motor load or dry mass flow rate fed to the refiners is to be controlled.
However, considering operating constraints, such approach is difficult to implement.
In this regard, in accordance with an embodiment of the proposed control approach, in addition to the feedback chip discharging control or instead of manipulating dosing screws 14, 15, 16 in a feedback mode, the CMS measurements can be used to predict the chips properties entering the refiners to perform a predictive control by manipulating the speed of screw feeder 27 of the primary refiner 31 to maintain the mass flow rate of dry materials, close to a setpoint. This second control scheme is also illustrated in Fig. 1, and in more detail in the schematic block diagram of Fig.
1B, involving a second control unit designated at 58, which integrates two mathematical models. The first model 34 is built to estimate the lag time between the CMS measurements and the primary refiner in order to predict the properties of chips arriving from main conveyer 22 at the primary refiner, and the second model 46 is built to estimate the current 'mass flow rate to the primary refiner 31 with respect to the screw feeder's speed. The control unit 58 then calculates the appropriate setpoint for the screw feeder's speed with respect to the required production rate. In other words, in the context of the present embodiment, the proposed approach consists of a predictive open loop control that employs two models that can be respectively called Preheater Washing Retention Steaming (PWRS) model referring to the steps of the first processing stage 17 of Fig.1, and Screw Feeder model as show in Fig.

1B. The PWRS model 34 estimates the delay time between the CMS 54 and the screw feeder 27 of the primary refiner 31 using process variables such as chips level in pre-heater 18, retention bin 23 and steaming vessel 24 as well as the production rate, to predict WCP entering the process at each time instant. The screw feeder model 46 represents the mass flow rate of the screw with respect to its rotational speed, and calculates the required speed for the screw considering the production rate and the WCP to keep the dry mass flow rate fed to the primary refiner at a constant value. As stated above, the second control unit 58 is especially suitable for mills in which chips classification is not performed and raw materials are stored in one pile. However, it should be understood that the second control scheme can be implemented both in mills with or without chips classification. As will be explained below in detail, the proposed control scheme involves in the first place measuring the delay time and identifying PWRS model parameters, considering the fact that TMP is a close process from the pre-heater to the latency chest with limited access to the materials. Consequently, no online re-tuning of the model parameters to adapt to new system's dynamic is generally possible. Furthermore, chips mixing and segregation in the bins and washing units of the first processing stage 17 shown in Fig. 1 may reduce the precision of the PWRS model prediction since the First-In-First-Out assumption is not completely valid anymore, as will be explained later in more detail.
In accordance with another embodiment, the prediction of the above models together with the process operating parameters can also be used to develop a third model to predict the quality of the pulp produced. The prediction can then be used to adjust the operating point of the process either by the process operators or through a third, model-based control unit designated at 60 in Fig. 1, which will be described later in more detail with reference to Fig. 1C. This other control scheme may involve the setpoint for SEC to stabilize CSF, optimal energy split between refiners or the required bleaching agent for optimal brightness of the pulp using the controlled pump 62.
As stated before, the general objective of the proposed approach is to reduce process variability to achieve energy efficiency and quality/productivity enhancement, which objective may be reached using on-line real-time chips properties measurements. As illustrated in Fig.1, the physical position of the CMS 54 is above the main conveyor after the chips piles and before the pre-heater bin 18 considering that chip properties measurements would otherwise be affected by the following processing. Such relative position has the inconvenient that there would be a lag time between the measurements of the CMS 54 and the arrival of the processed chips to the primary refiner 31. In this regard, a model can be developed to estimate the lag time between the CMS measurements and the arrival of chips to the primary refiner 31, hereinafter called 'chips model". Furthermore, a model can be developed for the screw feeder to control the dry-mass flow rate of the wood chips using calctAated lag time by the chips model applied to the CMS measurements, hereinafter called "screw model". Moreover, a model can be developed to predict pulp quality using the CMS
measurements and operating parameters, hereinafter called "pulp model'. This model would be used as the backbone of an ultimate Decision-Making-Support-System capable of taking necessary actions to control the quality of the pulp produced before its measurement by the PQM 57. The pulp model can also be employed to partially validate the chips and screw models. The scheme of the proposed control strategy is illustrated in Fig. 2, wherein pulp freeness (CSF) and fiber length (FL) are considered as quality indexes. It is to be understood that the same control approach may consider other pulp quality indexes, e.g.
brightness.
The phases of modeling, identification and validation of the various proposed models will now be described in detail. As to the chips model, one of the objectives of this phase is to determine the lag time between the CMS measurements and the arrival of the wood chips to the primary refiner. The model then predicts chips properties fed to the primary refiner at each instant using the CMS
measurements and applying calculated lag time. As stated above, this model is required when the positioning of the CMS in the process with its installation just before the primary refiner is not physically possible. The modeling procedure includes the five following steps:
Determining model inputs/outputs and other important factors;
Defining model structure;
Designing experiments and preparing test worksheet for model parameters identification;
Performing tests, refining test results, and optimizing model parameters;
Predicting chips characteristics arriving to the primary refiners; and finally =
Validating the model.
As for three first ones, the framework of well known Design of Experiment is followed, such as explained by Kettaneh-Wold et al. in "Design of Experiments" Umetrics Academy, 2000. As example, a list of identified factors is given in Table 1, Table 2, and Table 3.
Table 1: Chips recipe Description Abbr. Unit Type Use Range Black spruce HS Qualitativc,-MIxture Partially Controllab 60400 Fir Fl Qualitative-Mixture Partially Controllable 040 Aspen AS QualitatIve-Mixture Partially Controllable 0-5 Table 2: Chips properties Dmic r I pt ion A bbr. Unit Typo Use Range - Brightness Br Qualitative-Proems Uncontrollable 0-150 Moisture content MC % Qualitative-Proms Uncontrollable 20-50 Bulk density BuD kg/m3 Qualitaive-Prtxxsts Uncontrollable 150-250 Bask density Ban kg/m3 Qualitative-Process Uncontrollable 330430 Table 3: Process variables flasrlptLn Abbr. Unit Typo Use Range Pre-Heater Level % Qualitative-Process Controllable 20-100 Retention Bin Level ReL % Qualitative-Proem' Controllable 20-100 Steaming Vessel Line 1 Level iirL1 % Qualitative-Primes Controllable 30-70 Plug Screw Feeder Lino 1 % Qualitative-Frocses Controllable 0-100 Production Line 1 /We Prodl lid Qualitative-Proem Controllable 0-270 Steaming Vessel Line 2 Level StL2 % Qualitative-Process Controllsble 30-70 Plug Screw Feeder Line 2 Speed 9P2 % QuslitatIve-Proceet Controllable 0-100 Production Line 2 Rate Prod2 t/d Qualitative-Proems Controllable 0-270 As to model inputs/outputs, the list is divided into three categories, namely chips recipe, chips properties observed by the CMS, arid process variables. A
priori, chips recipe, which may be controllable to some extent, and properties/characteristics do not affect the lag time. However, since direct measurement of the lag time is not possible, the chips recipe is changed in the input to observe the impact on the refiner's motor load and disc gap of the primary refiner 31 to calculate the lag time indirectly. In which concerns the time lag response, it is defined, as mentioned above and in view of Fig. 2, as the time distance between the moment that the chips properties are measured by the CMS 54 and the instant that they enter the primary refiner through the screw feeder 27.
To perform modeling, the first step is to choose the model structure. Since the system elements are known, a physical modeling approach may be conveniently applied, on the basis of the simplified system schematically illustrated in Fig. 3. The basis on which a model structure can be proposed will now be described in detail.
The structure of the chips model may be inspired from the system of water tanks.
The difference is that contrary to fluid that is mixed after entering the tanks, solids are laid in horizontal layers. Such modeling system is illustrated in Fig. 4 for two tanks in series. The thickness of each layer is determined by the input flow rate GM
together with the time distance between each two measurements. For instance, the thickness of the first layer in the first reservoir is calculated from:
= rk+i) .t,õ
Ai Jr(k) (2) where L Tll is the layer thickness and tO) and i(ic+i) are two immediate consequent measuring instants with time distance rk = 0+1) t(10. In Equation (2), A1 is the surface of the tank 1. The equation can easily be generalized for any layer in both tanks 1 and 2 in Fig. 4. The objective is then to develop a mathematical model to calculate the arrival time of each layer, e.g. Layer 11, that has already passed from point AA to point BB. Point AA can be considered as a measurement point, Le.
in the case of this example, the CMS 54 and point BR, the consumption point i.e. at entry of the primary refiner 31. To model the lag time, the trip between points AA and BB is divided into two parts for each tank: the time that one layer reaches the bottom-line of the tank from the level at which it is formed, represented by Tu and Ty for the first and the second tank respectively, and the time that the layer passes from the bottom-tine of the first tank to the top of the other one, i.e. Tuand T.
The modeling system as introduced above may be further defined according to steady-state and dynamic cases as will now be explained in detail. If the system is in steady-state, chips flow rates are equal and invariant through time, that is:
Q = Q0(0= Qi(t) =Q20) (3) On the other hand, chips levels stay invariant. Hence, for any rzwe have:
7,11 = AiLu 9 (4) T21 A21,21 (5) where V and V, are chips volumes below Layers II and 21 of levels Ln and L21 respectively. For simplicity reasons, the transfer times from a tank to the following one, i.e. 712, are supposed to be constant. Production rate is also linked to the chips flow rate that is:
Pral = p x (6) where p is the chips bulk density. The total delay time is then calculated from:
Teak, = T" +r'2 + 1121 +Ta2 (7) Replacing Tu and T2I in the above Equation (7) by expressions in Equations (4) and

(5), we obtain:
Ttotai = + T12 4. A2Q42 ri (8) or Zeta/ = AlpTc--.0d + A01,7,14214 + 7121 +T

For:
= Alp X2 2-.7" A2p x3 = T12 + T22 (9) we have:
Twatp+,!017iXi+ '4271071X2+ X3 (10) In practice, production rate is controlled through the screw feeder's speed.
The relationship is supposed to be linear in the following form:
Prod=pQmpx0xSpzicaxSp (11) where a is the correction coefficient between chips flow rate and screw feeder's speed.
Hence:
Ttotar aLx1sp1 + aLx2sp1 x2 + x3 (12) The structure proposed in Equation (12) can easily be extended to a system with more than two tanks by adding any required terms. In which concerns the dimension of model parameters xõ two first parameters are [nej x flrghng = fh-ghni that represents the mass distribution in horizontal axis while the last parameter is in time N. Having test points, parameters are adjusted with respect to a predetermined criterion using any appropriate optimization method.
If the system is in dynamic state, the chips flow rates are not constant through time, and the steady-state method is not valid anymore. In this case, the following solution is proposed. For simplicity reasons, the focus is on Layer /1 as represented in Fig.4. The results can then be generalized to any other layer. The evolution of the volume of chips beneath Layer // is expressed by:

vtiTit = Qi(t)cit Jo (13) Since the analytical expression of Q/(0 is not known a priori, Equation (13) cannot be solved for ril. The intermediate solution is to suppose that during each Tk, flow rates are constant (a hypothesis that is classically used in sampling systems), that is:
V11 E Q11:71 (14) However, Equation (14) is not still easy to be implemented since the flow rate and volume are not directly measurable. It is supposed in Equation (11) that the flow rate and screw speed are linearly related, that is:
Ch(t) = a x Sp(t) (15) By supposing that flow rates are constant during each Tk, we have:
x SP1k (16) in discrete time. On the other hand, the initial volume beneath Layer ii while it is forming is:
Vu = Agin (17) Substituting Equation (16) and (17) in Equation (14) and bringing all the terms in one side, we have:
AiLit E aiSpark Al (18) Reformulating Equation (18), we have:
n A r :watt =O
L'd aiSPIk if=1 (19) To determine when Layer ii reached the bottom of the tank 1, the level of Layer /1 after passing through point AA and while it is forming is measured. Then, for each time stamp rb the Equation (19) is tested. If it is equals (or almost equals) to zero, it means that Layer 11 has already reached the bottom line and is leaving the tank 1.
Otherwise, the procedure is continued. As soon as Layer II reaches the bottom of the first tank, it would take TI2 instants to arrive to the next tank with a level of Layer 21. The above procedure is repeated to determine the time instant the layer passes point BB. The proposed concept can easily be generalized for several tanks in series as shown in Fig. 4.
Applying in the modeling system according to the dynamic case to develop the chips model, in view of Fig. 3, the following model structure is proposed:
D3'0 = (zi + x ReL) x ___ x Spi + 42 x 5p2 (20) MIA = 2:4 x StLl x __________ x Spl +
DrIA = za x 61M 1 x ca x Sp2+ xis and finally:
LJTI = Dlb + D71,2 Dr2 = +1,12õ2 (21) where x, are model parameters, and aj the correction coefficients between the production rate and the screw feeders speed:
Prod' = ai X Spa j=1,2 (22) Considering the proposed model structure and the objective of the experiment to model the lag time between the CMS and the primary refiner, a composite design is suggested to derive model parameters. A preliminary design worksheet using a 0-Optimal algorithm to minimize required experiments in a TMP mill provided with pulping process lines is presented in Table 4. 31 tests with 4 nominal operating points have been chosen between 460 possible choices. Due to the absence of hysteresis, test order is not important.
Table 4: preliminary experiments worksheet Exp. Run Chips No, Order Mixture Pr1,1361 ReL[N &LI(%) Elplf%) StL2N Sp2(9µ1 3 27 + 20 70 70 100 ao 0 16 + 70 20 TO 0 70 0

6 19 . 20 70 70 o 71) 0

7 12 + 20 20 70 100 70 0

8 33 _ 20 20 30 0 30 100

9 1 + 20 20 70 0 70 100 11 5 + 20 70 70 100 70 100 12 3 - 20 TO 30 0 30 33.33 13 . 12 + 20 70 30 100 43,33 100 14 34 - 20 70 30 66.66 79 0 33 + 20 53.33 70 0 30 100 16 4 = 70 20 30 100 56.68 0 17 6 + 70 20 30 66.66 30 100 18 29 - 70 20 70 0 39 66.66 19 9 + 70 20 56.66 100 70 100

10 - 70 70 30 100 70 33.33 21 7 4- 70 70 70 0 43.23 100 22 21 70 70 70 32.22 90 0 23 15 + 70 70 43.33 0 70 0 24 8 = 70 70 43,32 100 30 100 2 + 70 53.33 30 0 30 0 26 24 = 70 36.66 30 0 70 100 27 18 + 70 36,66 70 100 30 0 28 23 - 70 53.33 70 100 70 100 31 - 36.66 70 30 0 TO 100 31 30 + 53.33 70 30 100 30 0 32 22 _ 70 55 so so so 50 33 32 + TO 55 50 50 50 50 34 26 " 70 55 so so so 50

11 + 70 55 50 50 50 50 To measure the lag time, it is assumed that if any change is induced in the chips recipe, especially a change in chips bulk density, it would be possible to observe the impact of this change on one of the process measurements, that is in this case, the refiner's motor load. This hypothesis was approved by switching the alimentation of the process from a mixture of black spruce and fir to 100% of fir for a short time (approximately for 30 minutes) and then coming back to the original recipe. The variation of the dry bulk density and the impact on the motor loads of the primary refiners for line 1 and 2 are illustrated in the graphs of Fig. 5.
Considering that the lag time, i.e. response, is merely affected by some of the operating parameters such as chips level in the retention bin and steaming vessels and to minimize the test error, the number of the required variables was reduced, resulting in a lower number of tests to be performed. However, repeated tests were added to the worksheet to improve the modeling error characterization in the normal operating conditions. The performed test worksheet as well as the test results are provided in Table 5.

Table 5: Performed test worksheet and test results Tilt Lag RgL StLt StL2 Prod). Pricid2 Spl Sta Q utility No. :Uhl t (%) 91ET.-1 [t/41 [rid] [614 'Dm t 1.1 36 70 55.1 49.8 47.7 228.8 244.3 78.9 74j 100 1.2 30 70 55.1 40.8 47.7 228.8 244.3 78.9 74.1 100 2.1 x 269.7 244.4 02.9 74.2 0 3.1 94 67 54.9 40.2 38.1 246 242.5 83.5 74.3 100 3.2 90 65.7 54.9 40.2 38.1 246 242.5 83.5 74.3 100 4.1 20 59 54.95 40.2 38.3 243 2384 83.5 74.2 100 4.2 22 43.5 54.95 40.2 38.3 243 246.4 83.5 74.2 100 5.1 x 251.8 244.6 86.7 742 0 6.1 12 64 54.3 39.7 37.9 233.2 244.3 80.4 74.17 25 6.2 16 6/1 54.3 39.7 37,0 233-2 2443 89.4 74.17 26 6.3 30 6.6 54.3 39.7 37.9 233.2 244.3 80.4 74.17 50 6.4 30 6.6 54.3 39.7 37.9 233,2 244.3 80.4 74.17 60 7.1 28 9.4 14 26.4 26.5 2333 244.4 80.4 742 100 7.2 28 0.4 14 26.4 26.5 233.3 244.4 80.4 74.2 100 7.3 18 5.6 25 26.4 26.5 233.3 244.4 80.4 742 100 7.4 22 5.6 25 26.4 26.5 2333 244.4 80.4 742 100 8.1 18 8.2 24.0 39.9 38.1 233.7 244.6 80.4 74.2 75 6.2 16 8.2 24.0 30.0 38.1 233.7 244.6 80.4 74.2 TS
8.3 20 44.2 24.9 39.9 38.1 233.7 244.6 80.4 74.2 75 84 16 44.2 24.9 39.9 38,1 233,7 244.6 80.4 74.2 75 9.1 16 26.5 26 40.1 38.4 234.2 245.2 80.4 74.2 25 10.1 x 260,5 262 92.0 79.6 0 11.1 x 269,4 262.1 92.9 79,6 0

12.1 x 269.2 262.3 92,9 79.6 0

13.1 22 50 55.1 40 44.9 233.2 297 80.4 90.4 50 13.2 18 50 55.1 40 44.9 233,2 297 80.4 90.4 50 12.2 28 75 55.1 40 44.0 233.2 207 80.4 004 100 13.4 28 76 55.1 40 44.9 223,2 207 80.4 90.4 100

14.1 62 74.9 55 30.0 44.9 220,1 248.1 78.8 75.5 100 14.2 72 74.0 515 30.9 44,9 220,1 248.1 78.8 75.5 100 14.3 30 74.9 55 39.9 44.9 233.2 244.6 80.4 74.2 100 14.4 30 74.9 55 39.9 44.9 233,2 244.6 80.4 74.4 100

15.1 02 75 55 40.2 38.4 242.2 271.11 83.5 8/3 50

16.1 86 75.3 55.2 39.6 38,3 224 227.3 77.3 e8.8 60 16.2 86 73.3 55.2 39.6 38.3 224 227.3 77.3 613.8 50 16.3 82 75.3 55.2 39.6 38.3 224 227.3 '77.3 68.8 50 16.4 82 75_3 55.2 39.6 38.3 224 2273 77.3 68,8 50

17.1 68 73 55 40 38.4 260 261.7 89.8 79.5 100 17.2 60 TS 55 40 38.4 260 261.7 89.8 79.5 100 17.3 66 75 55 40 38.4 260 261.7 89.8 79.5 100 17.4 TO 75 55 40 38.4 260 261.7 89.8 70.5 100 From the data in Table 5, correction coefficients in Equation 22 that as calculated are:
a 2.90 a2 = 3.30 The model parameters are also identified by positive Least-Mean-Square method using the test results after transforming time scale from minute to second.
The nature of the model imposes that all parameters should be positive. The identified parameters are:
Xi = 5532, x2 = 1407, x3 = 67112, x4 = 0 while modeling error:
X5 = Pr T P
is characterized as a probability distribution in the following form:
P = [0.0808 0.4911 2.0562 5.9863 12.2990 18.300820.8686 20.7066 21.5480 24.7432 27.1538 25.1371 19.1455 12.75308.0659 41100 2.2364 0.7787 0.1892 0.03141 T 1-3.8988 3.4580 3.0180 2.6790 2.1300 - 1.6091 - 1.2691 -0.8192 - 0.3792 0.0608 0.5007 0.9407 1.3806 1.8206 2.2605 2.7005 3.1405 3.6804 4.0204 4.46031 x 1000 Considering the modeling error, the low quality prediction of the model is evident, which is especially due to unexpected chips mixing and segregation phenomena taking place in the bins, especially the pre-heater. This phenomenon can be affected by chips dimensions, moisture content and wood species. The mixing/segregation phenomenon should therefore be considered as an important factor since the first-in-first-out assumption for chips passing through the bins was not approved after performing and analyzing the tests. The significant modeling error should be handled in an appropriate way, and a solution is proposed in the prediction algorithm described below.
After modeling the lag time, the next step is to predict the chips properties arriving to the primary refiners. This is performed by a chips properties prediction algorithm, basically consisting of the following three steps:
1. accumulating CMS
measurements, batching the measurements every rk instants and calculating the mean value for each data batch;

2. calculating the lag time for each data batch; and 3. estimating the arrival time of the batches and predicting chips properties as fed to the primary refiner.
The details are provided in which follows. As to CMS data accumulation and batching, the CMS measurements are sent in every time-stamp to an appropriate plug-in provided in the second control unit, or as an alternative, can be sent to any appropriate data archiving module. Considering that punctual signals are subject to different types of perturbation/noise, and that it is useless to follow every individual measurement produced by the CMS, data are accumulated for a defined period and are batched while the mean value of the data represents the batched chips properties. The data batching and splitting is performed under one of the following conditions: minimum data batching period is fixed to a minimum splitting time;

maximum data batching period is fixed to a maximum splitting time; or data are split after a minimum splitting time and before a maximum splitting time if the standard deviation of the treated process variable (TPV) reaches a maximum data standard deviation, wherein the TPV is chosen as dry bulk density (DBO) in the present example. A graph of the data accumulation and batching and averaging procedure for the chips moisture content is shown in Fig. 6.
Calculating the lag time for each data batch and consequently, the remained time can easily be performed in static mode explained above, from:
remained time = estimated lag time - elapsed time Nevertheless, due to the evolution of the production rate and eventual startups and shutdowns of either of the production lines, the remained time is mostly dynamic. The second complexity level to be taken into account is the modeling error. As for the dynamic time distance, two different operating modes are proposed: the production rate is supposed to be constant in a static mode and the remained time decreases constantly with respect to the control unit clock; or the evolution of the remained time for each batch is re-calculated with respect to the elapsed time, the production rate during the elapsed time, and the actual production rate. Furthermore, three solutions according to respective modes are proposed to take into account the modeling error:

1. in the first mode (1), no modeling error is considered and the predicted chips properties belong to the batch with less remained time, as shown in the graph of Fig.7(a);
2. in the second mode (2), error distribution is applied to the mid-point of every data batch, the error distribution being used to weight the batches with respect to their distance from the primary refiner, and the chips properties prediction is then calculated from:
p WiPt+:;741Pi i+1 wi (23) where A is chips properties in the Zgh data batch, as shown in the graph of Fig. 7(b);
and 3. the third mode (3) is similar to the second, except the fact that two wings of the error distribution are applied to two extremes of the data batch, as shown in the graph of Fig. 7(c) .
Old data batches are eliminated as soon as the respective chips batches are consumed in the refiner. In mode 1, data batch i is eliminated as soon as the absolute temporal distance of data batch 1 + / becomes less than of the P. In modes 2 and 3, it happens when remained-time for data weights is negative and the data batch is not involved in prediction procedure anymore (error distribution has already passed the point 0 in remained-time axis).
For purposes of model simulation, the proposed algorithm was programmed.
The chips properties arriving to the primary refiner was estimated off-line using different operating modes and error handling methods. A batch of raw data with the sampling rate of Is was imported from the CMS database and was used to perform the simulation. Operating mode was set to 2 and the error handling method to 3.
Parameters and error distribution identified above was used to calculate the DEM
entering the primary refiners. Maximum and minimum splitting times were set to s and 560 s respectively. Maximum standard deviation was also set to 5 kgIrn3.
To simulate the variation of the production rate, it was reduced to its half value during samples 40 000 and 80000. The simulation results representing the prediction of DBD in the primary refiners, are illustrated in the graph of Fig. 8.

A brief discussion on the choices and the effect of different operating modes and error handling methods on the model prediction results is provided in which follows. As stated before, the evolution of the production rate can affect considerably the estimated arrival time of the chips batches to the refiners. The production rate is generally kept constant for a long period. However, if the production is subject to major changes in short intervals, it is suggested to use the second operating mode in dynamic state especially when the screw feeder is controlled automatically.
The manipulation of the screw feeder speed affects the volumetric consumption and consequently, affects the initial calculated arrival time. Chip moisture contents at constant and variable production rates are simulated, i.e. predicted, and compared in the graph of Fig. 9. Production rate is reduced by half between instances 2 000 and 4 000. The predicted values are shifted in time to fit the best to the original signal.
As stated above, the inconvenient of the first error handling mode (1) is that the mixing/segregation phenomenon of the chips along the way to the refiners from the CMS is ignored, being based on the first-in-first-out assumption that was not approved after the tests performed. However, the properties evolution is piecewise and therefore more adapted to be used in discrete piecewise control of the screw feeder's speed.
In the second mode (2), the mixing/segregation phenomenon is taken into account. However, the algorithm's stability is sensible to the error distribution and the duration of the batched data. If the error distribution is shorter than the maximum splitting data, then there would be periods between the ending of one batch and the start of the next one during which all the weights are zero and consequently, no prediction is performed. This case is illustrated in the graph of Fig. 10.
As for the third mode (3), the advantage is that the mixing/segregation phenomenon is taken into account as a probability distribution of the modeling error.
The problem of the second method mentioned above is also solved since there is always an overlap between the weight functions. However, if the error distribution is large in time (it is eventually the case in regard to the identification data-set), then the algorithm is not sensitive to fast evolution of the chips properties which are thus filtered. As a result, the potential gain due to process stability may not completely be achievable. Simulation of the model prediction in all three modes is illustrated in the graph of Fig. 11.
As soon as the chips model parameters are identified and the algorithm is put in place, the prediction mechanism can be validated. For that, process and CMS
data of four month of operation were analyzed to identify periods with high DBD
variability.

The effect of DBD variation can then be observed on the process variables, namely refiners motor loads and disc gaps. A sample of validation curves chosen is shown in the graphs of Fig. 12. Error distribution of the lag time prediction for both identification and validation data-sets are also illustrated in the graph of Fig. 13.
Instead of regression graph, the error distribution is illustrated due to the fact that almost all of the identification and validation points represent the normal operation of the process with almost the same lag times. Therefore, regression graph would results in a circular cloud of points with no explicit direction whereas by the error distribution, modeling error in identification and validation data-sets can be compared.
As to the screw feeder model 46 as part of the second control unit 58 shown in Fig. 1B, it includes a mathematical representation for dry materials flow rate (DMFR) with respect to the plug screw feeder speed in the primary refiner. The model is used for contronthg the speed to stabilize wood chips DMFR fed to the primary refiners. Since the feeder is a volumetric screw, the DMFR is a function of the rotational speed of the screw, chips bulk density and moisture content As to the model structure, it can be bunt from an estimation of the current production rate of the process. When chips on-line real-time density and moisture content measurements were not available, the production rate was traditionally estimated as a linear function of the screw speed, i.e. from Equation 22 above:
Prodi = x Sp;
(24) The coefficient a, is optimized so that the Equation 24 performs the best estimation of the production rate for a given screw speed. The drawback is that since chips bulk density and moisture content evolve through time, the dry mass flow rate and consequently the production rate also change. This evolution is not taken into account in the above representation. To improve the precision, the mean values are replaced by the current values through adding the time index. The value of a at any instant t is then calculated as:
DBD(t) aift) = 1=3 (25) Where DBD(0 and MTh are respectively chips dry bulk density fed into the screw feeder at instant t and the mean value of the latter over a long period of time. DBM0 is obtained from the prediction of the chips model. Consequently:
Prodi(t) = x DriD(t) ________________________________ Spi(t) D BD
(26) As to parameters identification, in Equation 26, two parameters are to be identified to calculate the current production rate that are a, and DDE. For the former, data in Table 5 was used to calculate a, that results in the following numerical values:
Fir = 2.90 = 3.30 The difference in the numerical values of the coefficients comes from different configurations of the screw feeders used with the control system of the mill.
As for the mean value of dry bulk density, the CMS bulk density and moisture content measurements over 6 months of operation were analyzed. The obtained mean value was estimated to be:
DID = 1.80.9387Ekg/m31 Knowing the above numerical values, the production rate of either of the lines can be calculated at any time instant t with respect to the current value of DBD
using Equation (26).
The current production rate is calculated using the prediction of the chips model of DBD(t) fed into the primary refiner and the current screw speed. However, the objective is to stabilize the DMFR by means of a screw feeder set-point. By solving Equation (26) for screw speed, we have:
Prod,ref4(t) RD 13 SPIV =
DBD
(27) Using the provided speed set-point, dry mass flow rate is stabilized and that would result in process variability attenuation.
For purposes of model validation, a concrete mathematical procedure can be hardly proposed to validate Equation (27). Nevertheless, if the reference speed is used, the variability attenuation of the process can be measured through motor load, disc gap and CSF stabilization. This variability attenuation can then be considered as an indirect validation of the screw model. Equation (27) is principally based on the assumption that the flow rate in the screw feeder is linear with respect to the screw speed. Although it is not true for the entire operating range of the feeder, it can be considered linear around a given point that is, in the case of the present example, a normal operating point of the process. The speed reference follows the evolution of chips bulk density and moisture content that are small compared to their mean value and consequently, the validity conditions are satisfied.
Optionally, as mentioned above and turning back to Fig.1, a bleaching agent model can be implemented in the third control unit 60, to optimize the consumption of the bleaching agent in the latency chest 35. Likewise the screw model, the bleaching agent model may use the prediction of the chips model on the brightness of the wood chips fed to the process. However, contrary to the screw model that intervenes at the screw feeders level, the bleaching agent model calculates the optimum amount of the bleaching agent considering the chips brightness evolution through time while respecting the required paper brightness constraints. It then provides a set-point for the control system of the mill commanding pump 6210 adjust the respective flow rate of the agent fed to the latency chest 35. The bleaching agent model may be built based on a principle quite identical to the one used for the screw model as described above. A more detailed schematic block diagram of the third control unit 60 is illustrated in Fig. 1C, which performs a predictive control of the pulp quality. By the quality in this example, we mean the brightness of the pulp in the latency chest 35. It is to be understood that other quality indicative parameters such as CSF and FL can be considered as well to be able to perform advanced model-based panel control of the TMP. Since the PQM 57 provides online pulp quality measurements, the model can be re-tuned in real-time. The proposed strategy consists of decomposing the pulp quality model into a model representing the TMP process (TMP Model) as well as the latency chest (LC model), represented by TMP/LC model 36 in Fig. 1C, and an adaptive module 39 receiving the output data from the pulp quality model, which can simply be a linear filter. The pulp quality model is preceded by a model of the first process stage (PWRS model) receiving CMS measurements and designated at 34.

An adaptive module optimizer 41 is provided for calculating the optimal adaptive parameters using the PQM measurements to enhance the quality prediction. This separation permits to use more complex/intelligent structures such as Neural Networks or knowledge-based models to represent the TMP/LC part which can not necessarily be re-tuned online in real-time. A controller 44 calculates the optimal bleaching agent flow rate to respect the minimum pulp brightness.
The pulp model is the last model and the most challenging one from the theoretical and technical points of view, considering that it is the pulp model that would be used to predict the quality of the pulp produced. In the context of the present example, quality is measured by the freeness (CSF) and the fibre length (FL). The main complexity dimensions in modeling the pulp quality for a TMP
process are:
1. High number of factors, i.e. process parameters affecting the pulp quality;
2. The needs for on-line real-time measurements of the raw materials; and 3. Lack of real-time measurements of the pulp quality.
As for the first dimension, process parameters can be classified into different classes; they can be controllable (directly or indirectly) or non-controllable, and measurable or non-measurable. Examples of controllable parameters are disc gap and dilution water flow rates in the refiner (direct control) or consistency (indirect control through dilution water flow rate), whereas chips properties fed into the refiners, e.g. bulk density and moisture content, are non-controllable parameters.
Refiners motor loads and the respective load split are measurable parameters while the mean residence time of wood chips in the refiner, which is used to calculate the refining intensity and freeness, are not measurable parameters. Non-measurable parameters could ultimately be estimated using physical models of the process sub-systems, on the basis of studies of the dynamic of the process.
As for the second dimension, an appropriate on-line instrument such as CMS
must be used to measure chips properties in real-time. Otherwise, the models are mainly based on the mean-value of chips properties over a long period of time and do not take into account the impact of the variation of raw materials on the pulp quality. Although interesting and promoting in academic, the direct consequence and the drawback of these models is the lack of precision in industrial scale and in operation.

As mentioned above in view of Fig. 1, pulp quality measurements are generally performed in the latency chest 35 after the last refining stage 33.
The latency chest is a reservoir that accommodates typically more than one hour of the line production mainly for fibre relaxation after the intense refining process and pulp homogenization as an intermediate stage before that pulp is being used in paper machines. In other words, the latency chest acts like a low pass filter on the pulp quality indexes; it eliminates high frequency evolution and smoothes the indexes toward long term mean values. This results in a low informative data-set using process historical data for model identification in high-frequency operating range, and consequently, low modeling quality in these conditions. In which follows, a first attempt to model the process is presented, using one of the existing physical models disclosed in the literature Then an empirical model is proposed, due to the lack of the precision of the physical model. In parallel, solutions are proposed for the above issues and are validated using process data.
Before starting the modeling, identification data-set should be prepared and conceived. This includes input/output variable selection, data collection, data alignment, and finally data rectification. Regarding Input/Output variable selection, as mentioned before, many factors affect the pulp quality in a TMP. However, there are a number of them that are more critical. To these, the wood chips properties should also be added according to the present invention, which is an improvement as compared to the heretofore prevailing models. A list of the main factors is provided in Table "6. In the present example, output variables are freeness and fibre length measured by the PQM 57.

Table 6: Pulp model input variables No. Typo Source Description Used Chips Property Chips Model Bask Density Yes 9 Chips Property Chips Model Flulk Density Yes 3 Chips Property Chips Model Dried Bulk Density Yes Chips Property Chips Model Moisture Content Yes Process Parameter DCS System Plug Screw Speed Primary Refiner Yes 6 Promos Parameter DCS System Plug Screw Speed Secondary Refiner No 7 Process Parameter DCS System Plug Screw Speed Tertiary Refiner No Process Parameter DCS System Production Rate Yes 9 Process Parameter DCS System Motor Load Primary Refiner Yes Process Parameter DCS System Motor Load Secondary Refiner Yes 11 Process Parameter DCS System Motor Load Tertiary Refiner Yes 12 Process Parameter DCS System Dilution Water Flow Rate Primary Yes Refiner Screw Feeder 13 Process Parameter DCS System Dilution Water Flow Rate Primary Yes Refiner - Flat Zone 14 Process Parameter DCS System Dilution Water Flow Rate Primary Yes Refloat. - Conical Zone Process Parameter DCS System Dilution Water Flow Rate Secondary Yes Refiner - Screw Feeder 16 Process Parameter DCS System Dilution Water Flow Rate Secondary Yes P.efiner - Flat Zone 17 Process Parameter DCS System Dilution Water Flow Rate Secondary Yes Refiner - C'onical Zone

18 Process Parameter DCS System Dilution Water Flow Rate Tertiary Yes Refiner - Screw Feeder

19 Process Parameter DCS System Dilution Water Flow Rate Tertiary Yes Refiner - Flat Zone Process Parameter DCS System Dilution Water Flow Rate Tertiary Yes Refiner - Conical Zone 21 PrOCOSS Parameter DCS System Disc Gap Primary Refiner- Flat No Zone 22 Process Parameter DCS System Disc Cap Secondary Refiner - Flat No Zone 23 Process Parameter DCS System Disc Gap Thrtiary Refiner- Flat No Zone 24 Process Parameter DCS System Disc Gap Primary Refiner - Conical Yet;
Zone 45 Process Parameter DCS System Dine Gap Secondary Refiner Conical Yes Zone

20 Process Parameter DCS System Disc Gap Tertiary Refiner - Conical Yes Zone 27 Process Parameter .DCS System Pulp Consistence Primary Refiner No 28 Process Parameter DCS System Pulp Consistence Secondary Refiner No 29 Process Parameter DCS System Pulp Consistence Tertiary Rdiner No Process Parameter DCS System Disc Age Primary Refiner No 31 Process Parameter DCS System Disc Age Secondary Refiner No 32 Process Parameter DCS System Disc Age Tertiary Refiner No As to data collection, several months of process and chips data were collected from the archives of the process information (PI) system of the mill. Details are provided in Table 7. The difference between the periods of chips model and process data is due to their availability and quality. This raw data are the grains of identification data-set used in both physical and empirical modeling of the pulp model.
Table 7: Pulp model identification data collection periods Type Source From To Sampling Rate Proii;:s 2ot0-05.01 2011 -03.01 30 mint itPs Input (hips Model 2010-06-30 2011-01-12 2 minutes Out put PQM 2u10-0T1-01. 2011-03.01 30 nUnutus The needs for data alignment and re-sampling will know be discussed. Chips properties and process data are already aligned with the required precision considering the fact that the chips model predicts the chips properties fed to the TMP. However, the measured pulp quality by the PQM has a lag time with respect to the other variables. As mentioned before, the latency chest is a reservoir that contains typically about 70 minutes of the pulp production. Ideally, it can be considered as a continuous stirred tank in which input/output flow rates are time-variants due to the production rate of the TMP and the consumption of paper machines. The PQM 57 samples the pulp after the latency chest 35 and before it enters the screens 37, 38. In this regard, two solutions are explored to align process and chips data with PQM measurements. The diagram of the input/output flows in the latency chest is illustrated in Fig. 14. For latency chest modeling purposes, a first approach considers the dynamic of the chest and the evolution of its state through time. It consists of developing an observer to estimate the pulp quality entering the latency chest from the measured values leaving it. In view of Fig. 14, for the output flow rate we have:
F0(t) = F4(t) -F3() (28) that can be calculated from measured values F3(t) and F4(t). To calculate F(t) and F/(0, the evolution of the volume in the chest is characterized in the following form:
(t) = Fi(t) F0(t) (29) where V(t) is the volume of pulp in the chest Knowing that V(t) = A x LW, the above equation is reformulated as:

ao Adi,(t) dt = Fo(t) (30) where A is the cross-section area and LO is the pulp level in the chest.
Discretizing it, we get:
A
¨T x [L(k + 1) ¨ 1,(k)] = Fi(k) Fo(k) (31) wherein T is the sampling rate. Fi(k) is then calculated in a recursive form from:
A
F(k) x [L(k +1) ¨ Mk)] + Fo(k) (32) Having been calculated the input and output flow rates, the system dynamics, as described by M.J. Willis. Continuous stirred tank reactor models" Department of Chemical and Process Engineering, University of Newcastle, March 2000, (http://lorien.ncl.ac.uk/mina/controlkien/cstr.pdf), can be defined by:
_____________________ = Fi(t)X;(t) ¨F0(t)X(t) dt (33) where X(t) is a pulp quality index either csF(t) or FL(t). Discrifizing the above equations, substituting v(o= A X 1,(0 and solving the equation for )4(k) we obtain:

X1(k) (A[X(k 1).L(k + 1)¨ X(k)1,(k)]
+ Fo(k)X(k)) Fi(k) (34) 1. AIX (k + 1)L(k +1) ¨ MOW)] Fo(k)X(k) Xi(k) = T Fi(k) F(k) (35) Either of the Equations (34) or (35) can be used to calculate the pulp quality indexes entering the latency chest.
Supposing that the system does not evolve through time, i.e. the volume of pulp is relatively constant most of the times in the chest, the lag time is also constant.
It would be equal to approximately three times of the time constant of the latency chest as reported by Willis cited above. The time constant r = 77 is calculated from the Laplace transform of Equation (33) representing system dynamics. Knowing that in a stable system, input and output flow rates as well as the pulp volume in the chest are constants, i.e. Fial= = F and Vit) = V, we have:
VX(S)S = FiX,(S) FOX(S) (36) that gives:
X(S)= Trv9+1 _VS) (37) for which r = F. As in the present example, if the reserve of the chest is of 77 min of the production that equals to 3 000 m3, the time constant is r 2: 77 min.
In practice, the level and consequently the volume in the chest are lower. For instance, if the level is 80% of its maximum, the time constant of the system is reduced to 62 min.
To validate either of the approaches, the partial least square (PLS) multivariate analysis was applied. Input and output vectors in this analysis were the input and output variables in the identification data-set. The analysis was performed for: 0; 30; 60; 90; 120; 150; 180; 210; 240; 270; and 300 min of lag time between the input (process data and chips model prediction) and output (PQM measurements) vectors. The analysis was also carried out for the input vector and the latency chest (LC) model estimation of the pulp quality entering the chest as the output vector.
From the percentage of variance explained by the PLS model with respect to the number of components , the best performance for CSF in both lines was obtained for:
30; 60; 90 min lag time. The same result was observed for fibre length in line 1 whilst for fibre length in line 2, 60; 90; and 120 min lag time produced the best PLS

modeling performance. Should it be theoretically more precise, numerical results demonstrated that using LC model do not result in more precision.
Consequently, to keep the modeling procedure as simple as possible and to avoid numerical instability that may occur using LC model in real-time operation, lag time between the process variables and PQM measurements was conveniently fixed to 75 min. PQM
measurements are then aligned with the process data as well as chips model prediction. All data in identification data-set are finally re-sampled with 30 min sampling rate.

Having been collected, aligned and re-sampled, the data pass through the rectification procedure. This procedure consists of:
1. Identifying shutdowns and faulty-periods and eliminating respective data;
2. Detecting outliers and eliminating respective data;
3. Filtering high-frequency measurement noises and perturbations; and 4. Identifying relatively unstable operating periods of the process and eliminating respective data from the identification data-set.
Data rectification includes removing shut-downs and faulty periods, eliminating outliers, and filtering measurement noises. Different methods and algorithms have already been developed in this regard, such as reported by Mingfang et al. in 'An integral approach to dynamic data rectification' Computers &
Chemical Engineering, 24(2-7):749- 753, 2000; and by Ragot et al. in Data validation using orthogonal filters' Control Theory and Applications, 1EE
Proceedings D, 139(1):47- 52, Jan 1992. A three-step heuristic procedure that may be used is presented in which follows. To detect shutdown, faulty periods or any undesired periods to be excluded from the analysis, control variables are used. This implies simple or combinatory conditions. Examples are for instance:
1. If the refiners motor loads 1 and 2 are more than 80% of the nominal load, and the production rate is more than 90% of the nominal value, then the process status is normal;
2. If the refiners motor loads 1 and 2 are less than 1% of the nominal load, then the process status is shut-down;
3. If the refiners motor load 1 is less than 1% and the refiner's motor load 2 greater than 80% of the nominal load, then the process status is faulty-mode;
4. If the refiners motor loads 1 and 2 are more than 80% of the nominal load, and the production rate is less than 10% of the nominal value, then the process status is faulty-mode.
The identified periods are then eliminated from the initial process data-set and only normal periods are kept for the next step.
An outlier is an observation that is numerically distant from the rest of the data and deviates significantly from the majority of the observations, as defined by Barnett et al. in "Outliers in Statistical Data' John Wiley & Sons Ltd., 3gi edition, 1998.

Several sources can be identified to cause an observed point to be an outlier.
One situation that happens frequently in industrial cases is division by zero in calculated variables when the denominator falls to a small value. An example of such variables is specific energy consumption (SEC) defined as the amount of energy required to produce a unit of product, mathematically stated as follows:
= Energy Consumption SEC _____________ Production Rate (38) During start-ups and shut-downs, or when the production is being stocked, the measurement of the production rate may be a very small value or even zero that causes division-by-zeros and consequently, produces outliers for instantaneous calculated value of the SEC. This can even cause stability problems in the control of the process if the control is performed using the latter without considering the validity conditions of the calculated value. Another source of outliers to be stated is the environmental noise and disturbances such as electromagnetic fields on measuring instruments. Several statistical techniques exist to detect outliers both for univariate and multivariate data. For univariate data, Chauvenet's criterion (Chauvenet "A
manual of spherical and practical astronomy. J.B. Lippincott & Co., 1863.), Peirce's criterion (Peirce On peirce's criterion" Proceedings of the American Academy of Arts and Sciences, volume 13, pages 348- 351, 1877), and Dixon's Q test (Rorabadier "Statistical treatment for rejection of deviant values: critical values of dixon's "Q"
parameter and related subrange ratios at the 95% confidence level" Analytical Chemistry, 63(2):139-146, 1991) can be stated. An online method for outliers' detection is also presented by Liu et al. in "On-line outlier detection and data cleaning" Computers & Chemical Engineering, 28(9):1635 - 647, 2004. As for multivariate data, Hotelling T-Square distance based on Principle Component Analysis (PCA) is a well-known technique (Kettaneh-Wold et at. " Multi and Megavariate Data Analysis: Principles and Applications, Umehics Academy, 2001) The Hotelling T-square for observation i (observation i in multivariate analysis X, is the vector of n variables observed simultaneously at instant I based on A
latent components is defined as:
A a c7-02 Aqs (39) where S'3,õ is the variance of ',according to the class model, iE T is the score matrix in PCA, and > A (N2 ¨ 1) N(N - A) (40) with the F distributed with A and N - A degrees of freedom, if 44(2_i) _________________________ X T2 Feriticat(P)A) (41) then observation i is outside confidence region (1 -p) of the model or in other words, it is an outlier. This technique is employed to detect and eliminate outliers from the process data.
To perform data filtering, likewise outlier detection techniques, different filtering methods have already been exploited to attenuate or eliminate measuring noise. The filtering can be performed frequency-based or time-based. For example, a fifth order Butterworth low-pass filter can be used to eliminate high-frequency measuring noise. The filter is used in its non-causal mode to avoid introducing inherent filtering lag time between filtered and non-filtered data.
Unstable periods are periods in which each of the main process variables variability is relatively high. This is especially due to the fact that the PQM signals are typically the moving average of 5 last measurements to damp measuring error.
Consequently, instantaneous PQM measurements in unstable periods are less reliable.
As stated before, variability for a time series x is defined as the standard deviation over the mean value of the series. Analyzed variables in the present example are freeness, fibre length, production rate, and reject-line load. If these variables are relatively stable, we can conclude that the process is in a stable period.
The stability condition can be defined as a threshold on variability.
Numerical results of stability analysis are provided in Table 8.
Table 8: Numerical results of the stable periods in the identification data-set Unstable Unstable Thistabki¨Unstatile 'Movable Line =Window Stability Periods Periods Periods Periods Periods Size Threshold Reject-Linn Production CSF FL Total Lotu I ['Xi Ravi%) rgi 6 ito mg 0.1 36,53 29.77 6.04 - 0.49 43.060 2 5 hours 0.1 40.15 29.55 5.73 0,28 43,89 The variability of studied variables with respect to the stability threshold in the identification data-set of the present example is illustrated in Fig. 15 and Fig. 16.
From 9409 points, 7169 are marked as outlier/unstable and are eliminated from the original data-set. Pulp model identification was then pursued using remaining points.
For purposes of empirical modeling, due to its flexibility and its ability to approximate any non-linear behavior as a universal approximator, neural network (NN) may be used in to model pulp quality. However, it is to be understood that any other appropriate modeling platform may also be used. Artificial NN have received considerable attention as a powerful tool to map and model any arbitrary continuous bounded nonlinear system, function or time series. In this regard, two major types of NN can be considered that are: Feed-Forward Neural Networks (FFNN) including multi-layer perceptron (MLP) and radial-basis function (RBF) networks; and Recurrent Neural Networks (RN N). Feed-Forward Neural Networks have proved extremely successful in pattern recognition problems and non-linear mapping in static systems while Recurrent Networks have been used in . associative memories, optimization problems and non-linear dynamic feedback systems. It is worth stating that from a system theoretic point of view, systems are either static or dynamic. By a static system, we mean a system that its response depends on the current inputs whereas in a dynamic system, it depends on both the current inputs and the previous state of the system in the past A MLP-NN with two hidden layers is illustrated in Fig.
17. In layer i, each inputs is weighted by an element in matrix W. The sum of the weighted inputs forms the inputs to the transfer function F. The output of each transfer function is then the input of the next layer. Transfer functions in a MLP-NN
are generally chosen identical. A list of potential functions is provided in Table 3.11.
Table 11: Common transfer functions in MLP neural networks Del,cription Ai action ii;;;I'jisfUnction loggig(u) Tan-Sigitioisi transfer function Flu) = tansig(u) Linear trraigfcr function F(u) purciiti(u) The identification is then defined in this context as optimizing the weight matrices so that the network approximates the best the target output. To train a MLFF-NN, the most well-known method is the back-propagation (BB). A general version of BB
is explained in detail by Narendra et al. in identication and control of dynamical systems using neural networks' IEEE Transactions on Neural Networks, 1(1):4-27, mar 1990. Based on that original work, extensive research has been done in this field and more efficient algorithms have been proposed among which the method presented by McLoone et at. in "A hybrid linear/nonlinear training algorithm for feedforward neural networks' IEEE Transactions on Neural Networks, 9(4):669-684, jul.1998, can be stated as an example. A generalized weight adaptation algorithm is also proposed by Kosmatopoulos et al. in "High-order neural network structures for identification of dynamical systems" IEEE Transactions on Neural Networks, 6(2):422-431, mar. 1995. However, a concern is to determine the optimal number of hidden layers as well as the number of nodes in each layer.
RBF-NN is an alternative form of feed-forward networks that has only two hidden layers. The transfer function of the first layer is a continuous non-linear function that calculates the distance of input vector from a given center and of the second layer is a linear line. A RBF-NN is illustrated in Fig. 18. Some of potential mathematical forms of transfer functions for RBF-NN are provided in Table 3.12 where c is the center of RBF and parameter a is the radius or width of e , as presented by Lian et at. in "Self-organizing radial basis function network for real-time approximation of continuous-time dynamical systems' IEEE Transactions on Neural Networks, 19(3):460-474, March 2008.
Table 12: Common transfer functions in RBF neural networks Description Function Gaussian Radial transfer function Ehti) ( "42-4:1 + COS( )) ¨ 4 <
Raised-Cosine Radial transfer function e(u) elsewhere Linear transfer function = P4w4in0 A complete theoretical discussion on the construction of RBF is presented by Karayiannis et al. in "On the construction and training of reformulated radial basis function neural networks' IEEE Transactions on Neural Networks, 14(4):835-846, July 2003, and requisite conditions for RBF are scrutinized by Huang et al. in "Universal approximation using incremental constructive feedforward networks with random hidden nodes" IEEE Transactions on Neural Networks, 17(4):879-892, July 2006. Similar to MLP neural networks architecture, the RBF-NN can model any continuous bounded non-linear mapping, but has the advantages of simpler topological structure and faster training that facilitates its use in adaptive situations.
Likewise MPL, the identification of RBF-NN basically consists of optimizing the weight matrices of the RBF. However, it can eventually include calculating optimal values of centers and the number of neurons in radial-basis hidden layer.
Due to its simplicity, a diversity of different approaches exists for RBF-NN
training. A hybrid training algorithm combining gradient-based optimization of non-linear weights with singular value decomposition computation of linear weights in one routine is proposed by Mason et al. in "Predictive control of a mixing tank using radial basis function networks" Decision and Control, Proceedings of the 35th IEEE, vol. 1, pp. 478-479, Dec 1996. In Liang et al. 'A fast and accurate online sequential learning algorithm for feedforward networks" IEEE Transactions on Neural Networks,17(6):1411 -1423, Nov. 2006, an algorithm called on-line sequential extreme learning machine (OS-ELM) is presented. The training data can be fed one-by-one or chunk-by-chunk that increases extremely learning speed compared to gradient-based methods. In Huang et al. cited above, it is proved that in order to let RBF-NN to work as universal approximator, it is simply required to randomly choose hidden nodes and then to adjust the output weights linking the hidden layers and the output layer. Based on this theory, the parameters of hidden layer are randomly selected and the output weights are analytically determined. The only parameter to be predefined is the number of hidden nodes. The performance of OS-ELM is evaluated against other training methods e.g. back-propagation and support vector machine by Huang et al. in "Extreme learning machine: Theory and applications"

Neurocomputing, 70(1-3):489-501, 2006. To improve the stability of the network, an ensemble of on-line sequential extreme learning machine (E0S-ELM) based on OS-ELM is proposed by Lan et al. in "Ensemble of online sequential extreme learning machine" Neurocomputing, 72(13-15):3391-3395, 2009. To handle data uncertainty and to approximate complex, continuous and deterministic systems from incomplete information, the mutual information radial basis function network MI-RBFNN is proposed by Deignan et al. in 'The MI-RBFN mapping for generalization"
Proceedings of the 2002 American Control Conference, volume 5, pages 3840-3845, vol.5, 2002, as an efficient, general and integrated method. A concern in feed-forward NN is how to determine the number of nodes in the hidden layer. In this regard, a growing and pruning strategy is proposed by Huang et al. in "A
generalized growing and pruning RBF (GGAP-RBF) neural network for function approximation' IEEE Transactions on Neural Networks, 16(1):57-67, Jan. 2005 that, based on the required modeling precision, adds or eliminates neurons. Weight matrices are also optimized in parallel after each iteration. An extension of this sequential learning algorithm is presented in Huan et al. (July 2006) cited above. An alternative sequential learning algorithm for function approximation and time-series prediction is also proposed by Zhihong et al. in An adaptive tracking controller using neural networks for a class of nonlinear systems" IEEE Transactions on Neural Networks, 9(5):947-955, Sep 1998, which results in minimal topology for the RBF-NN. That work combines the resource-allocation network growth criterion with a pruning strategy based on the relative contribution of each hidden unit to the overall network output. Two other relative growing/pruning strategies are presented in Lian et al.
cited above and in Feng et al. "Error minimized extreme learning machine with growth of hidden nodes and incremental learning" IEEE Transactions on Neural Networks, 20(8):1352-1357, Aug. 2009.
So as to model non-linear dynamic systems, RNN can be used. One of the first RNN is Hopfield neural model as described in Hopfield "Neural networks and physical systems with emergent collective computational abilities' Proceedings of the National Academy of Sciences of the United States of America, 79(8):2554-2558, April 1982, and Hopfield "Neurons with graded response have collective computational properties like those of two-state neurons' Proceedings of the National Academy of Sciences, 81:3088-3092, 1984. A typical RNN topology with feedback paths is illustrated in Fig. 19. Due to its specific topology, gradient-based algorithms are prominently used in parameter identification as reported by Narendra et al. cited above. Likewise FFNN, network identification means to optimize weight matrices.
The topology of the network is generally predetermined and is fixed during the identification.
As for identification strategies, a method for Hopfield network identification of time-varying dynamical systems is presented by Chu et al. in "Neural networks for system identification" Control Systems Magazine, IEEE, 10(3):31-35, April 1990, using least-squared estimation. A gradient-based algorithm is also proposed for general RNN by Narendra et al. cited above. A modified version of back-propagation for RNN is proposed by Williams et al. in "A learning algorithm for continually running fully recurrent neural networks" Neural Comput., 1:270-280, June 1989, which is based on the idea of unfolding RNN into a MLFF-NN that grows by one layer in each time step. The identification of Elma net, one of the simplest types of RNN is investigated by Pham et al. in "Identification of linear and nonlinear dynamic systems using recurrent neural networks' Artificial Intelligence in Engineering, 8(1):67-75, 1993. Approximation and learning properties of high-order recurrent NN is scrutinized in Kosmatopouios et al. cited above. The selected algorithm is then used to model dynamical systems. In Bailer-Jones et al. "A recurrent neural network for modelling dynamical systems" Network, 9(4531-547, 1998, RNN is identified and employed as a state observer.
As stated above, NN are able to approximate any arbitrary linear or non-linear function and time-series. It is because of this property that NN-based controllers have widely been developed for the compensation of the effect of nonlinearity and system uncertainties in control systems so that the system performance such as stability, convergence and robustness can be improved. In Narendra et al. " Neural networks in control systems' Proceedings of the 31st IEEE Conference on Decision and Control, pages 1-6 vol.1, 1992, both MLP-NN and RBF-NN are studied to address control of non-linear dynamical systems while gradient method is used to adjust their parameters. In Xiaohong et al. " Fuzzy adaptive control based on RBFN's Proceedings of the 1998 American Control Conference, vol. 3, pp. 1422-1426, June 1998, a fuzzy adaptive control strategy based on RBF-NN is presented for highly non-linear and multi operating mode plants.
Real-time control using NN is studied and scrutinized by Narendra in "Neural networks for real-time control' Proceedings of the 36th IEEE Conference on Decision and Control, vol. 2, pp. 1026-1031, Dec 1997. A control strategy for adaptive control of a class of non-linear discrete-time dynamical systems using NN is proposed by Chen et al. intelligent control using neural networks and multiple models., 2002, Proceedings of the 41st IEEE Conference on Decision and Control, vol. 2, pages 1357-1362, Dec. 2002. An intelligent two-level controller schema using NN is proposed by Narendra et al. in intelligent control using neural networks' Control Systems, IEEE, 12(2)1 1-18, April 1992, to control complex dynamic systems, wherein in the higher level of the schema, system failure is detected using pattern recognition technique and activates a stabilizing controller while in the second level, an adaptive controller is used to optimize the stabilized system. In Zhihong et al.
cited above, a RBF-NN-based adaptive tracking control scheme is proposed for a class of non-linear systems. The NN adaptively learns the bounds of uncertain dynamics and the output adaptively adjusts the gain of the sliding mode controller.
Two similar works in adaptive control are presented in Rovithalds et al.
"Adaptive control of unknown plants using dynamical neural networks. IEEE Transactions on Systems, Man and Cybernetics, 24(4400-412, March 1994, and Alanis et al.
'Discrete-time adaptive backstepping nonlinear control via high-order neural networks" IEEE Transactions on Neural Networks, 18(4)1 185-1195, July 2007. An issue in NN identification is incorporating prior knowledge into a NN. This issue is studied by Sartori et al. in Implementations of learning control systems using neural networks Control Systems, IEEE, 12(2):49-57, April 1992, wherein a particular NN
architecture and an associated design methodology are presented to store directly and systematically this knowledge into the NN without any training. Any appropriate NN such as disclosed in the above cited references may be used for purposes of empirical modeling for the needs of the proposed control approach.
Considering the above discussion and the dynamic of TMP process, FFNN is proposed to model pulp quality in the context of the present example. This is due to the fact that the system dynamics (except the latency chest) is static; the output of the process depends on the current operating parameters and the fed raw material properties while it is independent of the previous state of the TMP process.
In which follows, two FFNN that are MPL and RBF are exploited. The best configuration is obtained and the optimal parameters are identified. An adaptive strategy using a linear corrective module is also proposed to improve the modeling precision and eliminate systemic error.
Multi-Layer Perceptron Neural Network is first investigated for modeling. In the first step, the number of hidden layers is fixed to one. Apart from parameter identification, two issues are to be studied that are determining the optimal number of nodes in the hidden layer, and determining the Principal Component Analysis (PCA) coefficients.
Since the number of inputs is high, it is worth trying to reduce them to a lower number, using linear PCA such as taught in Fiori An experimental comparison of three PCA neural network? Neural Process. Lett.,11:209-218, June 2000; in Malthouse "Limitations of nonlinear PCA as performed with generic neural networks IEEE Transactions on Neural Networks, 9(1):165-173, Jan 1998; and in Liu et al. "A
hierarchical intrusion detection model based on the pca neural networks Neurocomputing, 70(7-9)1561-1568, 2007. In this regard, PCA coefficient means the percentage of the variation of the original input vector taken into account by latent variables. The coefficient varies between 0 and 1. For 1, all variation is taken into account, and the output of PCA is identical to the original input vector. If 0, no variation of the original vector is considered, which results in zero for all component coefficients. The number of latent variables is automatically determined in consequence. To obtain the optimal values for number of hidden nodes and PCA
coefficient, a 2D grid-based search experience was designed, the number of neurons changing from 10 to 200 with steps of 10 and PCA coefficient from 0.00 to 0.60 with steps of 0.20. At each node of this mesh, a MLP-NN was identified using 60% of the identification data-set for network training, 20% for validation to stop training and avoiding over-fitting and other 20% for an independent test. To test the stability of each configuration, this procedure was repeated 10 times. To save the space and avoid repetition of similar figures, only the performance results of the experience on freeness (CSF) of line 1 is illustrated in the graphs of Figs. 20 (in time domain), 21 (mean), 22 (standard deviation), 23 (best) and 24 (regression). The best performance modeling results for the other indexes are also provided in Table 3.13.
Table 13: Best performance of pulp quality modeling using MLP-NN
Neurons Regression Regmodon RegnNssion Regression Quality II id den PCA '11- ai ri in g Validation Test Total Index Layer Coeff. Data Data Data Data t. SI. 1 60 0.0 0,42 0.82 0.84 0.89 (M'2 50 0.6 0,95 0.82 0.83 0,41 I.'Ll 20 0.6 0,94 0.89 0.90 0,72 FL2 40 0.2 0.93 0.83 0.81 0.91 For purposes of Radial Basis Function Neural modeling, the same experience as described above was repeated. However this time, instead of exploring optimal number of neurons in the hidden layer, it is more interested to know the impact of radius of the transfer functions called hereafter the spread factor on the modeling performance. In this experience, the PCA coefficient was changed from 0.0 to 0.6 with the steps of 0.20 and spread from 0.01 to 10 with steps of 2. The number of neurons was determined automatically considering the trend of training and validation performance in each iteration. Best performance modeling results in term of CSF are provided in Table 14. Respective figures are illustrated in the graphs of Figs. 25 (time domain), 26 (mean), 27 (standard deviation), 28 (best) and 29 (regression).
Table 14: Best performance of pulp quality modeling using RBF-NN
Neurons Regression R egression Regression Regression Quality Hidden PCA Taiuing Val idat ion Test Total Index Layer Coeff. Data Data ____ Data Data __ CSF1 4.0 0.0 0.94 0.70 0.83 089 (.741.1 4.0 0.4 0.94 0.77 0.80 0 88 F1.1 4.0 0.4 0.97 0.79 0.89 0.92 n2 2.0 0.0 0.95 0.82 0,86 0.9/
Considering the results of the experiences, two different topologies can be proposed for freeness and fibre length. The mean value performance curves provide requisite information to determine optimal topology parameters, i.e. optimal number of neurons in hidden layer, PCA and spread coefficients. Performance standard deviation curves are used in parallel to determine the stability of NN
parameter identification in each configuration.
As to freeness, MLP-NN produces the best modeling performance. As an example, the number of neurons in the hidden layer is set to 150. PCA
coefficient is chosen to be one since it does not result in significant enhancement in the performance. A PCA coefficient of one reduces the complexity of the model as well.
As to fibre Length, RBF-NN is proposed for modeling with, for example, a spread coefficient of 4.0, a PCA coefficient set to one, and an optimal number of neurons in the hidden layer of 500 considering the trend of training and validation performances.
Using the above topologies, two models for freeness and two models for fibre length for line 1 and 2 were identified. Although the modeling results were satisfying, to take into account the variation of the process's dynamics through time, to cover the unidentified regions of the operating space, and to eliminate systemic errors that may occur during the operation of the model in real-time, an adaptive module can be employed in the design. In view of the general concept summarized above and illustrated in Fig. 2, an example of decomposition of the pulp model as well as the interaction with other elements will now be described in detail in view of Fig. 30. A
proposed adaptive module is composed of a linear corrector in the following form:
xeorrected(k) = a x xpredieted(k) + b (42) for x E [ CSF, FL). Adaptive parameters a and b are calculated such that the error between the measured and the corrected value be minimum, mathematically stated as:
[auptimal, boptinudi Mill E (xcorreeted(k--11)¨x,õ,,,,.õd(k-02 avb 6=1 (43) subject to:
a E forninl arnax]
E [Lbesp,UbcsF1 for x=CSF
E [Lb.pL Ub FL] for x=FL

where Ii is the optimization horizon, a parameter to be set by the user. The imposed conditions on a and b are applied to avoid bypassing of the model by the optimization algorithm in specific conditions. The proposed adaptive strategy was implemented and the predicted and corrected pulp quality indexes was compared to the measured ones using identification data-set. The distribution of the error for the predicted and corrected indexes is illustrated in Fig. 31 for an optimization horizon of 10 hours, a E
[0:9; 1:11 bcsFE [-50; 50] and bFLE [-1; 1]. The distribution curves suggest a minor enhancement on CSF and FL prediction. Adaptive parameters should be chosen with caution to improve prediction results. For instance, if the optimization is performed over a smaller horizon, e.g. 5 hours, prediction results are degraded as illustrated in Fig. 32. It should also be noted that in case of major modifications in the TMP such as changes in chips recipe or paper grade, the pulp model is to be re-trained for new conditions.

Claims (12)

1. A method for stabilizing dry-mass flow rate of wood chips to be fed from at least one discharging source to a chip refining line provided with a screw feeder at an entry thereof, through a first chip processing stage upstream said refining line, said method comprising the steps of:
i) estimating on-line a set of wood chip properties characterizing said wood chips prior to be fed to said first chip processing stage, to generate corresponding wood chip properties data, said set including moisture content and bulk density to derive from said wood chip properties initial dry mass flow rate data;
ii) feeding said initial dry mass flow rate data at an input of a predictive control model for deriving therefrom current speed setpoint data for said screw feeder according to a predetermined production rate of said refining line; and iii) manipulating the speed of the screw feeder in accordance with the setpoint data to substantially stabilize the dry mass flow rate of the wood chips fed at the entry of the refining line.

2. The method according to claim 1, wherein said predictive control model includes:
a first model representing said processing stage for estimating a lag time between said estimating step i) and the wood chips feeding at the entry of the refining line to generate predicted dry mass flow rate data; and a second model representing said screw feeder and receiving said predicted dry mass flow rate data and said predetermined production rate to derive therefrom said current speed setpoint data.

3. A method for stabilizing production of pulp from wood chips to be fed from at least one discharging source to a chip refining line provided with a screw feeder at an entry thereof, through a first chip processing stage upstream said refining line, said method comprising the steps of:
i) estimating on-line a set of wood chip properties characterizing said wood chips prior to be fed to said first chip processing stage, to generate corresponding wood chip properties data, said set including moisture content and bulk density to derive from said wood chip properties initial dry mass flow rate data, said set further including at least one reflectance-related property;
ii) feeding said initial dry mass flow rate data at an input of a first predictive control model for deriving therefrom current speed setpoint data for said screw feeder according to a predetermined production rate of said refining line;

iii) controlling the speed of the screw feeder in accordance with the setpoint data to substantially stabilize the dry mass flow rate of the wood chips fed at the entry of the refining line;
iv) feeding the wood chip properties data at an input of a further predictive control model for deriving therefrom further current setpoint data for an associated control parameter for controlling a quality parameter characterizing the pulp to produce from said wood chips; and v) manipulating the control parameter in accordance with the further setpoint data to substantially stabilize the pulp quality parameter.

4. The method according to claim 3, wherein said associated control parameter is bleaching agent flow rate, said pulp quality parameter being pulp brightness.

5. The method according to claim 3, wherein said first predictive control model includes:
a first model representing said processing stage for estimating a lag time between said estimating step i) and the wood chips feeding at the entry of the refining line to generate predicted dry mass flow rate data; and a second model representing said screw feeder and receiving said predicted dry mass flow rate data and said predetermined production rate to derive therefrom said current speed setpoint data.

6. The method according to claim 5, wherein said further predictive control model uses said first processing stage model for estimating the lag time between said estimating step i) and the wood chips feeding at the entry of the refining line to generate predicted wood chip properties from which said further current setpoint data are derived.

7. An apparatus for stabilizing dry-mass flow rate of wood chips to be fed from at least one discharging source to a chip refining line provided with a screw feeder at an entry thereof, through a first chip processing stage upstream said refining line, said apparatus comprising:
means for estimating on-line a set of wood chip properties characterizing said wood chips prior to be fed to said first chip processing stage, to generate corresponding wood chip properties data, said set including moisture content and bulk density to derive from said wood chip properties initial dry mass flow rate data, a control unit provided with a predictive control model receiving at an input thereof said initial dry mass flow rate data, for deriving therefrom current speed setpoint data for said screw feeder according to a predetermined production rate of said refining line; and means for manipulating the speed of the screw feeder in accordance with the setpoint data to substantially stabilize the dry mass flow rate of the wood chips fed at the entry of the refining line.

8. The apparatus according to claim 7, wherein said predictive control model includes:
a first model representing said processing stage for estimating a lag time between the estimation of said wood chip properties and the wood chips feeding at the entry of the refining line to generate predicted dry mass flow rate data; and a second model representing said screw feeder and receiving said predicted dry mass flow rate data and said predetermined production rate to derive therefrom said current speed setpoint data.

9. An apparatus for stabilizing production of pulp from wood chips to be fed from at least one discharging source to a chip refining line provided with a screw feeder at an entry thereof, through a first chip processing stage upstream said refining line, said apparatus comprising:
means estimating on-line a set of wood chip properties characterizing said wood chips prior to be fed to said first chip processing stage, to generate corresponding wood chip properties data, said set including moisture content and bulk density to derive from said wood chip properties initial dry mass flow rate data, said set further including at least one reflectance-related property;
a control unit provided with a predictive control model receiving at an input thereof said initial dry mass flow rate data, for deriving therefrom current speed setpoint data for said screw feeder according to a predetermined production rate of said refining line;
means for manipulating the speed of the screw feeder in accordance with the setpoint data to substantially stabilize the dry mass flow rate of the wood chips fed at the entry of the refining line;
a further control unit provided with a predictive control model receiving at an input thereof said wood chip properties data, for deriving therefrom current setpoint data for an associated control parameter for controlling a quality parameter characterizing the pulp to produce from said wood chips; and means for manipulating the control parameter in accordance with the further setpoint data to substantially stabilize the pulp quality parameter.

10. The apparatus according to claim 9, wherein said associated control parameter is bleaching agent flow rate, said pulp quality parameter being pulp brightness.

11. The apparatus according to claim 9, wherein said first predictive control model includes:
a first model representing said processing stage for estimating a lag time between the estimation of said wood chip properties and the wood chips feeding at the entry of the refining line to generate predicted dry mass flow rate data; and a second model representing said screw feeder and receiving said predicted dry mass flow rate data and said predetermined production rate to derive therefrom said current speed setpoint data.

12. The apparatus according to claim 11, wherein said further predictive control model uses said first processing stage model for estimating the lag time between the estimation of said wood chip properties and the wood chips feeding at the entry of the refining line to generate predicted wood chip properties from which said further current setpoint data are derived.