CN106931453B - The forecasting system and method for circulating fluid bed domestic garbage burning emission of NOx of boiler - Google Patents

Abstract

The invention discloses a kind of real-time estimate system and method for circulating fluid bed domestic garbage burning emission of NOx of boiler.Using the method for BP neural network algorithm and the particle swarm optimization algorithm integrated moulding on multiple populations for introducing Operator of Pattern Search, the system and method for constructing a kind of fast, economical and adaptive updates carry out real-time estimate to boiler smoke NOx emission, avoid the modelling by mechanism work of very complicated.The dynamic variation characteristic of NOx emission is characterized using the nonlinear dynamic characteristic of BP neural network algorithm, generalization ability and real-time estimate ability；Optimizing is carried out to the initial weight and threshold value of BP neural network using particle swarm optimization algorithm, reduces the possibility that BP neural network is absorbed in local optimum in the training process；Operator of Pattern Search and migration mechanism on multiple populations are introduced, improves the diversity and local search ability of particle swarm optimization algorithm solution, particle cluster algorithm optimizing is reduced and calculates the possibility for being absorbed in local optimum.

Description

The forecasting system and method for circulating fluid bed domestic garbage burning emission of NOx of boiler

Technical field

The present invention relates to energy project field, especially, is related to a kind of circulating fluid bed domestic garbage burning boiler NOx rows Put forecasting system and method.

Background technology

Volume reduction, minimizing, innoxious and recycling of the waste incineration due to can well realize technology of garbage disposal, closely In more than ten years, under the guiding of national related industry policy, domestic waste incineration industry achieves vigorous growth.From last century The nineties, domestic more scientific research structures are burnt to Municipal Solid Waste in China (Municipal Solid Waste, MSW) Mechanism has carried out a large amount of further investigations, and the burning for having grasped high, complicated component the domestic waste of mixed collection, moisture is special Property, the inferior fuel such as coal, gangue recirculating fluidized bed (Circulating Fluidized Bed, CFB) is burnt according to China On the basis of the development Experience of technology, refuse incinerator of circulating fluid bed is have developed, from Zhejiang University's exploitation in 1998 First fluidized bed refuse incinerator puts into operation beginning, show suitable for domestic high-moisture, calorific value is relatively low and fluctuates Property very big house refuse the characteristics of carrying out large-scale burning disposal.At present, CFB garbage incineration technologies at home more Individual city has carried out popularization and application, ends for the end of the year 2015, the built platform of garbage burning boiler more than 70 in the country, day processing quantity of refuse 6.9 ten thousand tons, important contribution is made that for the incineration treatment of garbage industry in China.

Boiler tail flue gas NOx discharge is one of important symbol for weighing the whether environmentally friendly operation of boiler,《House refuse is burnt Burn contamination control standard》The 1 small hourly value and 24 small hourly values of (GB 18485-2014) regulation NOx emission concentration must not distinguish More than 300mg/m³And 250mg/m³, otherwise will face the punishment of environmental administration's suspending operations for consolidation.Meanwhile producers and custodian Member is likely to require the emission behaviour of NOx under certain operating condition, in order to which the operation to boiler optimizes adjustment.Therefore, The NOx emission predictive model tool of one enough accuracy of structure is of great significance.

Researcher both domestic and external is studied the modeling of the NOx emission characteristic of CFBB, mainly have with Lower several method.It is a kind of be according to CFB boiler kinetics of combustion, hydrodynamics, heat and mass characteristic, by rational letter Change and established after assuming, mechanism model is established by way of mathematical description.This method can reflect the change of NOx discharge Trend, but as it is assumed that deviation between model and true model and be unable to reach enough accuracy；Another method be On the basis of substantial amounts of bench run or field test, established by the method for regression analysis and change spy on NOx emission The empirical model of property.This method needs to expend substantial amounts of manpower and materials, and time cost is high, while can not guarantee test covering institute Some operating modes, there is certain limitation；The third method utilizes the simplification of Fluid Mechanics Computation, numerical heat transfer and chemical reaction Mechanism simulation stove combustion process, accurately solve NOx generation situation, it is shown that there is good effect very big development to dive Power.But this method be primarily limited to fluid mechanic model and chemical reaction Reduced mechanisms and actual conditions gap, it is necessary to High-end allocation of computer and very long calculating time, therefore in this way still in the initial development stage.In addition, CFB The feeding system uniformity of garbage burning boiler is poor, and the hot-restriking die for entering stove rubbish is big, complicated components, polygons are strong, is One of main difficulty faced in NOx emission modeling process, the NOx emission characteristic model that it requires established have well Adaptive ability, above-mentioned three kinds of modeling methods have still been short of in this respect.

With the development of electronic technology, computer technology and information technology, Distributed Control System (Distributed Control System, DCS) running of CFB life burning boilers is widely used in, include temperature, pressure, flow etc. Process data including parameter is all completed and must preserved, and is that people recognize comprising abundant procedural information in these historical datas Know and understand one of important channel of production process, there is very high tap value, ground for the application of Intelligent data mining algorithm Study carefully and using providing excellent hardware and software platform.Error-duration model artificial neural network (Back Propagation Artificial Neural Network, BPANN) be a kind of typical multilayer perceptron (Multilayer Perceptron, MLP), a key character of neutral net is exactly the black box characteristic between input and output, and it will treat that modeling is considered as one Individual black box, it is indifferent to how complicated the internal mechanism of problem to be solved is, is only concerned the input and output of system.This causes nerve net Modeling of the network particularly suitable for the NOx emission characteristic of station boiler, can get around band in NOx generating process in this way The difficult points such as some hysteresis qualitys, non-linear and time variation, realize the complex mapping relation between NOx discharge and each influence factor. Meanwhile BP networks have self-learning capability, it can be trained according to fresh sample, the adaptive parameter for adjusting network.Work as coal When kind, dust stratification and equipment performance (such as performance change of garbage feeding system) change, NOx emission characteristic model also will Change therewith.NOx emission characteristic model can be adjusted with on-line training using the self-learning property of BP networks, ensure model Precision.

But the performance of BP networks has dependence to a certain extent to initial weight and threshold value, if initial weight and threshold value What is set is undesirable, and BP networks are easily trapped into local optimum.In order to improve this problem, this patent is introduced into Operator of Pattern Search Particle cluster algorithm on multiple populations be incorporated into BP networks, with it, come the initial weight to BP networks, then threshold value optimizes.Grain Swarm optimization is the Swarm Intelligent Algorithm simulated the predations such as flock of birds and grown up, and it utilizes individual right in colony The shared motion for making whole colony of information produces the evolutionary process from disorder to order in problem solving space, so as to obtain Optimal solution.This algorithm causes the attention of academia with the advantages that its realization is easy, precision is high, convergence is fast, and excellent in function Change field obtains a wide range of applications.But particle swarm optimization algorithm has certain blindness, repeatedly as random search algorithm It is less efficient for later stage Local Search.Therefore, using collaboratively searching pattern on multiple populations, it is total to by information between population Personal Enjoy, and the information interchange between population, the diversity of solution space is improved, the probability for being absorbed in local extremum can be reduced.Together When by simplex search operator be introduced into come, improve the algorithm later stage local search ability.

NOx emission predictive system has higher computational load, therefore, in order to improve NOx emission predictive system-computed effect Rate is, it is necessary to rationally set the frame structure of system.

The content of the invention

In view of the above-mentioned deficiencies in the prior art, it is an object of the present invention to provide a kind of circulating fluid bed domestic garbage burning boiler The forecasting system and method for NOx emission.The present invention is on the basis of CFB Domestic refuse incinerator operation mechanisms are analyzed, selection The input variable of NOx emission predictive model, modeling is trained to sample set using BP neural network, and utilizes and introduce simplex The particle cluster algorithm on multiple populations of operator optimizes to the initial weight and threshold value of BP neural network, finally utilizes Heterogeneous Computing ring Border builds NOx emission predictive system.

The technical solution adopted for the present invention to solve the technical problems is：A kind of circulating fluid bed domestic garbage burning boiler The real-time estimate system of NOx emission.The system and the Distributed Control System and production management system phase of CFBB Even, including data communication interface and host computer, are trained and more in host computer (server) to NOx emission predictive model Newly, the model trained is sent to Distributed Control System, production management system (client) by communication interface after allowing, it is described on Position machine includes：

Signal acquisition module.The module is being burned when specifying house refuse for gathering CFB Domestic refuse incinerators Operating condition state parameter and performance variable, and form refuse thermal value forecast model input variable training sample matrix X (m × N), m represents number of samples, and n represents the number of variable；

Data preprocessing module.Gross error processing and random crror processing are carried out to X (m × n), to forgo those not It is the deceptive information for reflecting boiler accidental conditions, the unusual service conditions such as boiler shutdown, banking fire, batcher blocking is excluded, In order to avoid the different harmful effects to caused by model performance of dimension between the parameter of forecast model and the order of magnitude, training sample Input variable is mapped to after normalized in [0,1] section, the training sample X after being standardized^*(m×n).In advance Processing procedure is carried out using following steps：

1.1) criterion is reached according to Rye, rejects the outlier in training sample X (m × n)；

1.2) boiler shutdown operating condition is rejected, the aperture of burner hearth feeder and batcher is zero during boiler shutdown, and Temperature is close to normal temperature in burner hearth；

1.3) burner hearth banking fire operation conditions is rejected, primary air fan, overfire air fan air-introduced machine burner hearth feeder during boiler pressuring fire Aperture with batcher is zero, but the temperature of burner hearth emulsion zone maintains 350 DEG C~450 DEG C；

1.4) reject batcher and block operating mode, batcher, which blocks, needs what operations staff was shot by the camera of material inlet Picture judges that feed situation when batcher blocks, operations staff can significantly heighten the aperture of batcher, be reflected in fortune In row data, i.e., the aperture of batcher is more than 35%；

1.5) data normalization processing.According to formula (1) by the section of data variable mappings to [0 1].

X in formula_JThe vector that J variables are formed is represented, min () represents minimum value, and max () represents maximum.

Expertise library module.Training sample is constantly updated using the method for rolling time window so that training sample begins Be maintained at eventually in newest state, rolling time window method refers to since current time, backtracking L (unit second) length when Between size.

Model modification determination module.The performance of current predictive model is detected, when relative prediction residual exceedes ± 5%, then Decision model needs to be updated.

Intelligent modeling module.Intelligent modeling module is the core of NOx emission predictive system, and the module is first with introducing The particle cluster algorithm on multiple populations of Operator of Pattern Search carries out optimizing to the initial weight and threshold value of BP neural network, then will obtain Optimal initial weight threshold value is assigned to BP neural network model, and is trained based on this.Algorithm steps are as follows：

2.1) initialization algorithm parameter.The parameter of BP neural network model and optimizing algorithm is configured, including BP god Implicit number of plies hl, node in hidden layer hn through network, training iterations gen1, learning rate η, hidden layer neuron activation letter Several classes of types；The maximum optimizing algebraically T of particle cluster algorithm on multiple populations_max, maximum inertia weight ω_max, minimum inertia weight ω_min, speed Degree renewal coefficients R₁、R₂、R₃、R₄, population quantity pop, the number of particles ind of single species；The factor alpha of simplex algorithm, tighten Coefficient θ, spreading coefficient γ, constriction coefficient β and search precision ε.

2.2) population is initialized.It is by the way of real coding, all weight thresholds of BP neural network model are orderly Encode among a particle, and each weight threshold is generated as to a real number between [0,1] at random.

2.3) individual extreme value and colony's extreme value are initialized.The initial weight and threshold value that are included in each particle are assigned to BP god Through network model, and combined training sample carries out learning training, and the forecast model obtained using training calculates NOx emission predictive valueBy predicted valueWith actual measured value y^*It is compared, and the fitness value of particle is used as using error sum of squares MSE Fitness, fitness calculation formula are as follows：

Initialization extreme value of the fitness value being calculated of each particle as the particle in itself, among each population Value minimum MSE is as colony's extreme value.

2.4) more new particle.According to newest individual extreme value and colony's extreme value, according to (3) formula and (4) formula more new particle Speed v_idAnd position x (t)_id(t)：

v_id(t+1)=ω v_id(t)+c₁r₁(p_id-x_id(t))+c₂r₂(p_gd-x_id(t)) (3)

x_id(t+1)=x_id(t)+v_id(t+1) (4)

In formula, t is the optimizing algebraically of particle swarm optimization algorithm, r₁, r₂It is the random number between [0,1], p_idRefer to I-th of particle is sought so far searches optimal location, p_gdRefer to the population and search for optimal location so far.Further, it is The defects of improvement basic particle group algorithm is easily trapped into local extremum and slow convergence rate, introduces on the basis of PSO algorithms Dynamic aceleration pulse c₁、c₂With inertia weight ω：

Wherein, T_maxFor maximum optimizing algebraically, ω_maxFor maximum inertia weight, ω_minFor minimum inertia weight, R₁、R₂、R₃、 R₄For constant.

2.5) particle fitness value calculation.The fitness value of particle after renewal is calculated according to formula (2).

2.6) more new individual extreme value and colony's extreme value.Using fitness value as evaluation index, more contemporary particle and previous generation Fitness value size between particle, if the fitness value of current particle is better than previous generation, the position of current particle is set Individual extreme value is set to, otherwise individual extreme value is kept constant.While the optimal particle of the present age all particle fitness values is obtained, and with Previous generation optimal particles are compared, if the fitness value of contemporary optimal particle is better than the fitness of previous generation optimal particles Value, then be arranged to global optimum by the adaptive optimal control angle value of contemporary particle, and otherwise global optimum is kept constant.

2.7) judge whether to need to carry out simplex search.A simplex search is carried out every 10 generations, is entered if desired Row simplex search then performs step 2.8), otherwise performs step 2.9).

2.8) simplex search.Simplex method (Simplex Method, SM) is a kind of traditional local search of algorithm, it Amount of calculation is small, search speed is fast, has very strong local search ability, can largely make up PSO local optimal searching energy The awkward situation of power relative deficiency.The method constructs a polyhedron with D+1 summit first in D dimension spaces, obtains each summit Adaptive value, and optimum point therein, secondary advantage and most not good enough are determined, then by the strategy such as reflecting, expanding, shrinking or compressing Find out one it is more better, substitution it is most not good enough, so as to form new polyhedron, such iteration can find or approach one most Advantage.Jacobi matrix and Hessian matrix of this method without solved function, without carrying out complicated matrix operation, have Very strong universality.Initial simplex { x is constructed first⁰,x¹,…,xⁱ,…,x^D, x⁰Searched for every sub- population optimal Solution, xⁱAccording to formula (8), (9) generation：

K=-0.05+0.1r (9)

J represents jth dimension variable in formula, and r is obeyed as equally distributed random number on [0,1].

The size of D+1 summit of simplex according to target function is renumberd, meets the numbering on summit：

fitness(x⁰)≤fitness(x¹)≤…≤fitness(xⁱ)≤…≤fitness(x^D) (10)

OrderIfThen stop iteration output x⁰。

2.9) immigrant's operation.In nature, the different population of a species is distributed in different regions, on the one hand each Groupy phase has facilitated their different existence to independently striving for resource from nature used in oneself under different regional conditions Pattern and evolution degree；On the other hand, possibly through migrating between each population, make to contact each other, reach mutual It is connected with nothing, learns from other's strong points to offset one's weaknesses, the purpose of common evolutionary.(Multi-Population, the MP) population on multiple populations that the present invention uses Algorithm is exactly this phenomenon for having used for reference generally existing in nature.Use each time in searching process, between population and unidirectionally follow The mode of ring migration carries out immigrant's operation, and the excellent individual in the 1st population moves to the 2nd, and the 2nd is moved to the 3rd, with such Push away, to the last one is moved to first.Migration rate P between population_i=0.04, i.e., P before ranking in expression source population_i× P after ranking among 100% individual replacement target population_i× 100% individual, the friendship of optimal knowledge between population is completed with this Stream.

2.10) algorithm stop condition judges.Judge whether to reach maximum iteration or reach the requirement of precision of prediction, The return to step 2.4 if being not reaching to), continued search for using the cluster radius of renewal, otherwise exit search, perform step 2.11)。

2.11) optimal weight threshold particle is exported.

2.12) the initial weight threshold value among optimal particle is assigned to BP neural network model, and combined training sample enters Row study.

2.13) precision of prediction of model is verified.The predicted value of model and actual value are contrasted, relative prediction is calculated and misses Difference.

2.14) relative prediction residual is judged whether within ± 5%, step 2.15) is performed if meeting to require, otherwise Return to step 2.12), the parameter of neural network model, and re -training again.

2.15) output meets desired NOx emission predictive model.

Communication module.The module sends the NOx emission predictive model for meeting to require to function Distributed Control System, production Management system.

A kind of circulating fluid bed domestic garbage burning emission of NOx of boiler Forecasting Methodology, this method comprise the following steps：

1) analyze circulating fluid bed domestic garbage burning boiler operation mechanism and NOx formation mechanisms, select rubbish to Doses, coal-supplying amount, primary air flow, secondary air flow, flue gas oxygen content, combustion chamber draft, bed temperature, burner hearth freeboard temperature conduct The input variable of NOx emission predictive model.

2) training sample is gathered.The historical data that input variable is gathered from database is spaced according to set time, or The operational factor under operating mode is specified in collection, forms the training sample matrix X (m × n) of NOx emission predictive mode input variable, m tables Show number of samples, n represents the number of variable, while gathers output training sample Y of the corresponding NOx discharge as model (m×1)；

3) data prediction.Gross error processing and random crror processing are carried out to X (m × n), those are not to forgo Reflect the deceptive information of boiler accidental conditions, the unusual service conditions such as boiler shutdown, banking fire, batcher blocking are excluded, are Avoid different to caused by the model performance harmful effects of dimension and the order of magnitude between the parameter of forecast model, training sample is defeated Enter variable to be mapped to after normalized in [0,1] section, the training sample X of the input variable after being standardized^* The training sample Y of (m × n) and output variable^*(m×1)。

4) intelligent algorithm integrated moulding.Calculated first with the population on multiple populations for introducing Operator of Pattern Search to BP neural network Initial weight and threshold value carry out optimizing, and obtained optimal initial weight threshold value then is assigned into BP neural network model, and with this Based on be trained.Algorithm steps are as follows：

4.1) initialization algorithm parameter.The parameter of BP neural network model and optimizing algorithm is set in the step Put, including the implicit number of plies hl of BP neural network, node in hidden layer hn, training iterations gen1, for learning rate η, implicit Layer neuron activation functions type；The maximum optimizing algebraically T of particle cluster algorithm on multiple populations_max, maximum inertia weight ω_max, it is minimum Inertia weight ω_min, speed renewal coefficients R₁、R₂、R₃、R₄, population quantity pop, the number of particles ind of single species；Simplex The factor alpha of algorithm, tighten coefficient θ, spreading coefficient γ, constriction coefficient β and search precision ε.

4.2) population is initialized.It is by the way of real coding, all weight thresholds of BP neural network model are orderly Encode among a particle, and each weight threshold is generated as to a real number between [0,1] at random.

4.3) individual extreme value and colony's extreme value are initialized.The initial weight and threshold value that are included in each particle are assigned to BP god Through network model, the forecast model obtained using training calculates NOx emission predictive valueBy predicted valueWith actual measured value y^* It is compared, and the fitness value fitness using error sum of squares MSE as particle, fitness calculation formula are as follows：

Initialization extreme value of the fitness value being calculated of each particle as the particle in itself, each population it is central Value minimum MSE is as colony's extreme value.

4.4) more new particle.According to newest individual extreme value and colony's extreme value, according to (2) formula and (3) formula more new particle Speed v_idAnd position x (t)_id(t)：

v_id(t+1)=ω v_id(t)+c₁r₁(p_id-x_id(t))+c₂r₂(p_gd-x_id(t)) (2)

x_id(t+1)=x_id(t)+v_id(t+1) (3)

In formula, t is the optimizing algebraically of particle swarm optimization algorithm, r₁, r₂It is the random number between [0,1], p_idRefer to I-th of particle is sought so far searches optimal location, p_gdRefer to the population and search for optimal location so far.Further, it is The defects of improvement basic particle group algorithm is easily trapped into local extremum and slow convergence rate, introduces on the basis of PSO algorithms Dynamic aceleration pulse c₁、c₂With inertia weight ω：

Wherein, T_maxFor maximum optimizing algebraically, ω_maxFor maximum inertia weight, ω_minFor minimum inertia weight, R₁、R₂、R₃、 R₄For constant.

4.5) particle fitness value calculation.The fitness value of particle after renewal is calculated according to formula (1).

4.6) more new individual extreme value and colony's extreme value.Using fitness value as evaluation index, more contemporary particle and previous generation Fitness value size between particle, if the fitness value of current particle is better than previous generation, the position of current particle is set Individual extreme value is set to, otherwise individual extreme value is kept constant.While the optimal particle of the present age all particle fitness values is obtained, and with Previous generation optimal particles are compared, if the fitness value of contemporary optimal particle is better than the fitness of previous generation optimal particles Value, then be arranged to global optimum by the adaptive optimal control angle value of contemporary particle, and otherwise global optimum is kept constant.

4.7) judge whether to need to carry out simplex search.A simplex search is carried out every 10 generations, is entered if desired Row simplex search then performs step 4.8), otherwise performs step 4.9).

4.8) simplex search.Simplex method (Simplex Method, SM) is a kind of traditional local search of algorithm, it Amount of calculation is small, search speed is fast, has very strong local search ability, can largely make up PSO local optimal searching energy The awkward situation of power relative deficiency.The method constructs a polyhedron with D+1 summit first in D dimension spaces, obtains each summit Adaptive value, and optimum point therein, secondary advantage and most not good enough are determined, then by the strategy such as reflecting, expanding, shrinking or compressing Find out one it is more better, substitution it is most not good enough, so as to form new polyhedron, such iteration can find or approach one most Advantage.Jacobi matrix and Hessian matrix of this method without solved function, without carrying out complicated matrix operation, have Very strong universality.Initial simplex { x is constructed first⁰,x¹,…,xⁱ,…,x^D, x⁰Searched for every sub- population optimal Solution, xⁱAccording to formula (7), (8) generation：

K=-0.05+0.1r (8)

J represents jth dimension variable in formula, and r is obeyed as equally distributed random number on [0,1].

The size of D+1 summit of simplex according to target function is renumberd, meets the numbering on summit：

fitness(x⁰)≤fitness(x¹)≤…≤fitness(xⁱ)≤…≤fitness(x^D) (9)

OrderIfThen stop iteration output x⁰。

4.9) immigrant's operation.In nature, the different population of a species is distributed in different regions, on the one hand each Groupy phase has facilitated their different existence to independently striving for resource from nature used in oneself under different regional conditions Pattern and evolution degree；On the other hand, possibly through migrating between each population, make to contact each other, reach mutual It is connected with nothing, learns from other's strong points to offset one's weaknesses, the purpose of common evolutionary.(Multi-Population, MP) the hereditary grain on multiple populations that this patent uses Swarm optimization is exactly this phenomenon for having used for reference generally existing in nature.Each time in searching process, using single between population Immigrant's operation is carried out to the mode of circulation migration, the excellent individual in the 1st population moves to the 2nd, and the 2nd is moved to the 3rd, with This analogizes, and to the last one is moved to first.Migration rate P between population_i=0.04, i.e., P before ranking in expression source population_i P after ranking among × 100% individual replacement target population_i× 100% individual, optimal knowledge between population is completed with this Exchange.

4.10) algorithm stop condition judges.Judge whether to reach maximum iteration or reach the requirement of precision of prediction, The return to step 4.4 if being not reaching to), continued search for using the cluster radius of renewal, otherwise exit search, perform step 4.11)。

4.11) optimal weight threshold particle is exported.

4.12) the initial weight threshold value among optimal particle is assigned to BP neural network model, and combined training sample enters Row study.

4.13) precision of prediction of model is verified.The predicted value of model and actual value are contrasted, relative prediction is calculated and misses Difference.

4.14) relative prediction residual is judged whether within ± 5%, step 2.15) is performed if meeting to require, otherwise Return to step 2.12), the parameter of neural network model, and re -training again.

4.15) output meets desired NOx emission predictive model.

5) model adaptation updates.When the error of NOx discharge and model prediction discharge capacity exceedes ± 5%, immediately more New model.

The beneficial effects of the invention are as follows：Gone through using the operation mechanism and operation of circulating fluid bed domestic garbage burning boiler It is excellent using BP neural network algorithm and the population on multiple populations for introducing Operator of Pattern Search in history data on the basis of tacit knowledge Change the method for algorithm integration modeling, construct the system and method for a kind of fast, economical and adaptive updates to boiler smoke NOx Discharge carries out real-time estimate, avoids the modelling by mechanism work of very complicated.Wherein, the non-linear of BP neural network algorithm is utilized Dynamics, generalization ability and real-time estimate ability characterize the dynamic variation characteristic of NOx emission, are operations staff and design The variation characteristic that personnel grasp understanding NOx emission provides new approach；Using particle swarm optimization algorithm to the first of BP neural network Beginning weights and threshold value carry out optimizing, reduce the possibility that BP neural network is absorbed in local optimum in the training process；Introduce more Population migration mechanism, the diversity of particle swarm optimization algorithm solution is improved, reduce particle cluster algorithm optimizing calculation and be absorbed in local optimum Possibility；Operator of Pattern Search is introduced, improves the local optimal searching ability of particle swarm optimization algorithm；The computing environment of isomery, is carried significantly High model construction efficiency, makes computer resource to have played the performance of maximum.Whole modeling process clear logic, auto-modeling Degree is high, is easy to grasp and promotes.Well-drilled NOx emission predictive model can be used for the actual motion for instructing operations staff Process, those System design based on model algorithms can be serviced, or it is mutually complementary with NOx hard ware measures system as soft measuring instrument Fill check.

Brief description of the drawings

Fig. 1 is the structure chart of system proposed by the invention.

Fig. 2 is the structure chart of master system proposed by the invention.

Fig. 3 is the system construction drawing that the present invention uses BP neural network model.

Fig. 4 is the flow chart of intelligent modeling method proposed by the invention.

Embodiment

The present invention is described in further detail with specific embodiment below in conjunction with the accompanying drawings.

Embodiment 1

Reference picture 1, Fig. 2, Fig. 3, Fig. 4, a kind of circulating fluid bed domestic garbage burning emission of NOx of boiler provided by the invention Forecasting system, including circulating fluid bed domestic garbage burning boiler, for the Distributed Control System of boiler operatiopn control, data Communication interface, database and host computer.Database reads data by data communication interface from Distributed Control System, is used in combination In the training study and test of host computer, host computer carries out data exchange, institute by data communication interface and Distributed Control System The host computer stated includes on-line study, online updating, verification portion and on-line prediction part.Specifically include：

Signal acquisition module.The module is being burned when specifying house refuse for gathering CFB Domestic refuse incinerators Operating condition state parameter and performance variable, and form refuse thermal value forecast model input variable training sample matrix X (m × N), m represents number of samples, and n represents the number of variable；

Data preprocessing module.Gross error processing and random crror processing are carried out to X (m × n), to forgo those not It is the deceptive information for reflecting boiler accidental conditions, the unusual service conditions such as boiler shutdown, banking fire, batcher blocking is excluded, In order to avoid the different harmful effects to caused by model performance of dimension between the parameter of forecast model and the order of magnitude, training sample Input variable is mapped to after normalized in [0,1] section, the training sample X after being standardized^*(m×n).In advance Processing procedure is carried out using following steps：

1.1) criterion is reached according to Rye, rejects the outlier in training sample X (m × n)；

1.2) boiler shutdown operating condition is rejected, the aperture of burner hearth feeder and batcher is zero during boiler shutdown, and Temperature is close to normal temperature in burner hearth；

1.3) burner hearth banking fire operation conditions is rejected, primary air fan, overfire air fan air-introduced machine burner hearth feeder during boiler pressuring fire Aperture with batcher is zero, but the temperature of burner hearth emulsion zone maintains 350 DEG C~450 DEG C；

1.4) reject batcher and block operating mode, batcher, which blocks, needs what operations staff was shot by the camera of material inlet Picture judges that feed situation when batcher blocks, operations staff can significantly heighten the aperture of batcher, be reflected in fortune In row data, i.e., the aperture of batcher is more than 35%；

1.5) data normalization is handled.According to formula (1) by the section of data variable mappings to [0 1].

X in formula_JThe vector that J variables are formed is represented, min () represents minimum value, and max () represents maximum.

Expertise library module.Training sample is constantly updated using the method for rolling time window so that training sample begins Be maintained at eventually in newest state, rolling time window method refers to since current time, backtracking L (unit second) length when Between size.

Model modification determination module.The performance of current predictive model is detected, when relative prediction residual exceedes ± 5%, then Decision model needs to be updated.

Intelligent modeling module.Intelligent modeling module is the core of NOx emission predictive system, and the module is first with introducing The population on multiple populations of Operator of Pattern Search is calculated carries out optimizing to the initial weight and threshold value of BP neural network, then will obtain most Excellent initial weight threshold value is assigned to BP neural network model, and is trained based on this.Algorithm steps are as follows：

Intelligent modeling module.Intelligent modeling module is the core of NOx emission predictive system, and the module is first with introducing The population on multiple populations of Operator of Pattern Search is calculated carries out optimizing to the initial weight and threshold value of BP neural network, then will obtain most Excellent initial weight threshold value is assigned to BP neural network model, and is trained based on this.Algorithm steps are as follows：

2.1) initialization algorithm parameter.The parameter of BP neural network model and optimizing algorithm is set in the step Put, make the implicit number of plies hl=1 of BP neural network, node in hidden layer hn=20, training iterations gen1=150, study Rate η=0.1, hidden layer neuron activation type function are tansig functions；The maximum optimizing algebraically of particle cluster algorithm on multiple populations T_max=100, maximum inertia weight ω_max=1.4, minimum inertia weight ω_min=0.4, speed renewal coefficients R₁=1, R₂= 0.5、R₃=6, R₄=2, population quantity pop=5, the number of particles ind=20 of single species；Factor alpha=1 of simplex algorithm, Tighten coefficient θ=0.5, spreading coefficient γ=2, constriction coefficient β=0.5 and search precision ε=0.0001.

2.2) population is initialized.It is by the way of real coding, all weight thresholds of BP neural network model are orderly Encode among a particle, and each weight threshold is generated as to a real number between [0,1] at random.

2.3) individual extreme value and colony's extreme value are initialized.The initial weight and threshold value that are included in each particle are assigned to BP god Through network model, and combined training sample carries out learning training, and the learning process of BP neural network model is as follows：

The key of BP algorithm includes from input layer to hidden layer again the information forward-propagating of output layer and error from output layer Arrive the back transfer of input layer, two process cycle alternations, untill error convergence again to hidden layer.Information from i-th layer just The output relation that is conveyed into m-th of neuron to jth layer is：

WhereinFor the output of m-th of neuron of jth layer, f (x) is activation primitive, and p represents of i-th layer of neuron Number,Represent nth iteration when i-th layer of k-th of neuron to the neuron weights,For nth iteration when i-th layer The output of k-th of neuron.

Feedback error process is the error energy energy according to desired output during nth iteration and real output value With using gradient descent method regulating networks weights, error function value is reached minimum.In order to improve convergence rate, often in network Momentum term is added in weighed value adjusting：

Network weight when middle w (n+1) is (n+1)th iteration, η are learning rate, and E is error energy function, α₀For momentum .

The forecast model obtained using training calculates NOx emission predictive valueBy predicted valueWith actual measured value y^*Carry out Compare, and the fitness value fitness using error sum of squares MSE as particle, fitness calculation formula are as follows：

Initialization extreme value of the fitness value being calculated of each particle as the particle in itself, each population it is central Value minimum MSE is as colony's extreme value.

2.4) more new particle.According to newest individual extreme value and colony's extreme value, according to (5) formula and (6) formula more new particle Speed v_idAnd position x (t)_id(t)：

v_id(t+1)=ω v_id(t)+c₁r₁(p_id-x_id(t))+c₂r₂(p_gd-x_id(t)) (5)

x_id(t+1)=x_id(t)+v_id(t+1) (6)

In formula, t is the optimizing algebraically of particle swarm optimization algorithm, r₁, r₂It is the random number between [0,1], p_idRefer to I-th of particle is sought so far searches optimal location, p_gdRefer to the population and search for optimal location so far.Further, it is The defects of improvement basic particle group algorithm is easily trapped into local extremum and slow convergence rate, introduces on the basis of PSO algorithms Dynamic aceleration pulse c₁、c₂With inertia weight ω：

Wherein, T_maxFor maximum optimizing algebraically, ω_maxFor maximum inertia weight, ω_minFor minimum inertia weight, R₁、R₂、R₃、 R₄For constant.

2.5) particle fitness value calculation.The fitness value of particle after renewal is calculated according to formula (2).

2.6) more new individual extreme value and colony's extreme value.Using fitness value as evaluation index, more contemporary particle and previous generation Fitness value size between particle, if the fitness value of current particle is better than previous generation, the position of current particle is set Individual extreme value is set to, otherwise individual extreme value is kept constant.While the optimal particle of the present age all particle fitness values is obtained, and with Previous generation optimal particles are compared, if the fitness value of contemporary optimal particle is better than the fitness of previous generation optimal particles Value, then be arranged to global optimum by the adaptive optimal control angle value of contemporary particle, and otherwise global optimum is kept constant.

2.7) judge whether to need to carry out simplex search.A simplex search is carried out every 10 generations, is entered if desired Row simplex search then performs step 2.8), otherwise performs step 2.9).

2.8) simplex search.Simplex method (Simplex Method, SM) is a kind of traditional local search of algorithm, it Amount of calculation is small, search speed is fast, has very strong local search ability, can largely make up PSO local optimal searching energy The awkward situation of power relative deficiency.The method constructs a polyhedron with D+1 summit first in D dimension spaces, obtains each summit Adaptive value, and optimum point therein, secondary advantage and most not good enough are determined, then by the strategy such as reflecting, expanding, shrinking or compressing Find out one it is more better, substitution it is most not good enough, so as to form new polyhedron, such iteration can find or approach one most Advantage.Jacobi matrix and Hessian matrix of this method without solved function, without carrying out complicated matrix operation, have Very strong universality.Initial simplex { x is constructed first⁰,x¹,…,xⁱ,…,x^D, x⁰Searched for every sub- population optimal Solution, xⁱAccording to formula (10), (11) generation：

K=-0.05+0.1r (11)

J represents jth dimension variable in formula, and r is obeyed as equally distributed random number on [0,1].

The size of D+1 summit of simplex according to target function is renumberd, meets the numbering on summit：

fitness(x⁰)≤fitness(x¹)≤…≤fitness(xⁱ)≤…≤fitness(x^D) (12)

OrderIfThen stop iteration output x⁰。

2.9) immigrant's operation.In nature, the different population of a species is distributed in different regions, on the one hand each Groupy phase has facilitated their different existence to independently striving for resource from nature used in oneself under different regional conditions Pattern and evolution degree；On the other hand, possibly through migrating between each population, make to contact each other, reach mutual It is connected with nothing, learns from other's strong points to offset one's weaknesses, the purpose of common evolutionary.(Multi-Population, MP) the hereditary grain on multiple populations that this patent uses Swarm optimization is exactly this phenomenon for having used for reference generally existing in nature.Each time in searching process, using single between population Immigrant's operation is carried out to the mode of circulation migration, the excellent individual in the 1st population moves to the 2nd, and the 2nd is moved to the 3rd, with This analogizes, and to the last one is moved to first.Migration rate P between population_i=0.04, i.e., P before ranking in expression source population_i P after ranking among × 100% individual replacement target population_i× 100% individual, optimal knowledge between population is completed with this Exchange.

2.10) algorithm stop condition judges.Judge whether to reach maximum iteration or reach the requirement of precision of prediction, The return to step 2.4 if being not reaching to), continued search for using the cluster radius of renewal, otherwise exit search, perform step 2.11)。

2.11) optimal weight threshold particle is exported.

2.12) the initial weight threshold value among optimal particle is assigned to BP neural network model, and combined training sample enters Row study.

2.13) precision of prediction of model is verified.The predicted value of model and actual value are contrasted, relative prediction is calculated and misses Difference.

2.14) relative prediction residual is judged whether within ± 5%, step 2.15) is performed if meeting to require, otherwise Return to step 2.12), the parameter of neural network model, and re -training again.

2.15) output meets desired NOx emission predictive model.

Communication module.The module sends the NOx emission predictive model for meeting to require to function Distributed Control System, production Management system.

Embodiment 2

Reference picture 1, Fig. 2, Fig. 3, Fig. 4, a kind of circulating fluid bed domestic garbage burning emission of NOx of boiler provided by the invention Forecasting Methodology, this method comprise the following steps：

1) analyze circulating fluid bed domestic garbage burning boiler operation mechanism and NOx formation mechanisms, select rubbish to Doses, coal-supplying amount, primary air flow, secondary air flow with, flue gas oxygen content, combustion chamber draft, bed temperature, burner hearth freeboard temperature make For the input variable of NOx emission predictive model.

Domestic domestic waste is mostly mixed collection, is caused into factory, to enter stove component of refuse complex, general main Including main components such as rubbish from cooking, paper, plastics, rubber, fabric, wood, bamboo and inorganic matters, low heat value, height are shown Moisture and the larger feature of fluctuation.In order to ensure the stable burning of refuse incinerator of circulating fluid bed, it will usually add coal As auxiliary fuel.Burning of the rubbish in recirculating fluidized bed is a sufficiently complex vigorous physical chemical change process, rubbish Rubbish can undergo several processes after burner hearth is entered：Dry heat, Volatile and burning, coke burning.It is light in rubbish Frangible component such as paper paper, plastics and fine grained can enter upper furnace in the presence of fluidized wind, and experience is dried, waved The precipitation of hair point and a series of processes such as burning and the burning of carbon residue；And density is larger, moisture content is high and particle size compared with The component that the big component such as terminal velocity such as wood, rubbish from cooking is more than fluidizing velocity can fall into emulsion zone, and in emulsion zone Heated, burn by bed, different from the heat release rule of coal, the component of rubbish higher moisture low heat value can inhale in emulsion zone Substantial amounts of heat is received, and substantial amounts of volatile matter burns in suspension section.

NOx discharge is by the thermo parameters method situation in burner hearth, oxygen concentration distribution situation and had in CFB tail flue gas What the distribution situation of machine volatile matter concentration determined.Coal-supplying amount, feeding coal and a secondary air flow have together decided on thermo parameters method, oxygen Gas concentration is distributed and organic volatile matter concentration distribution, and they are anti-by bed temperature, burner hearth freeboard temperature and flue gas oxygen content Mirror and.It particularly to be noted that in actual moving process, it may appear that the uneven situation of temperature field, component field distribution, and It can not be known completely by Oxygen Amount in Flue Gas measuring point and fire box temperature measuring point, and the fluctuation situation of combustion chamber draft can be to a certain degree Upper their fluctuation situation of reflection, therefore also using it as one of input variable of model.

2) training sample is gathered.The historical data that input variable is gathered from database is spaced according to set time, or The operational factor under operating mode is specified in collection, forms the training sample matrix X (m × n) of CO emitted smoke mode input variables, m tables Show number of samples, n represents the number of variable, while gathers output training sample Y of the corresponding CO discharge capacitys as model (m×1)；

3) data prediction.Gross error processing and random crror processing are carried out to X (m × n), those are not to forgo Reflect the deceptive information of boiler accidental conditions, the unusual service conditions such as boiler shutdown, banking fire, batcher blocking are excluded, are Avoid different to caused by the model performance harmful effects of dimension and the order of magnitude between the parameter of forecast model, training sample is defeated Enter variable to be mapped to after normalized in [0,1] section, the training sample X of the input variable after being standardized^* The training sample Y of (m × n) and output variable^*(m×1)。

4) intelligent algorithm integrated moulding.Calculated first with the population on multiple populations for introducing Operator of Pattern Search to BP neural network Initial weight and threshold value carry out optimizing, and obtained optimal initial weight threshold value then is assigned into BP neural network model, and with this Based on be trained.Algorithm steps are as follows：

4.1) initialization algorithm parameter.The parameter of BP neural network model and optimizing algorithm is set in the step Put, including the implicit number of plies hl of BP neural network, node in hidden layer hn, training iterations gen1, learning rate η, hidden layer Neuron activation functions type；The maximum optimizing algebraically T of particle cluster algorithm on multiple populations_max, maximum inertia weight ω_max, it is minimum used Property weights omega_min, speed renewal coefficients R₁、R₂、R₃、R₄, population quantity pop, single species number of particles ind,；Simplex is calculated The factor alpha of method, tighten coefficient θ, spreading coefficient γ, constriction coefficient β and search precision ε.

4.2) population is initialized.It is by the way of real coding, all weight thresholds of BP neural network model are orderly Encode among a particle, and each weight threshold is generated as to a real number between [0,1] at random.

4.3) individual extreme value and colony's extreme value are initialized.The initial weight and threshold value that are included in each particle are assigned to BP god Through network model, and combined training sample carries out learning training, and the learning process of BP neural network model is as follows：

The key of BP algorithm includes from input layer to hidden layer again the information forward-propagating of output layer and error from output layer Arrive the back transfer of input layer, two process cycle alternations, untill error convergence again to hidden layer.Information from i-th layer just The output relation that is conveyed into m-th of neuron to jth layer is：

WhereinFor the output of m-th of neuron of jth layer, f (x) is activation primitive, and p represents of i-th layer of neuron Number,Represent nth iteration when i-th layer of k-th of neuron to the neuron weights,For nth iteration when i-th layer The output of k-th of neuron.

Feedback error process is the error energy energy according to desired output during nth iteration and real output value With using gradient descent method regulating networks weights, error function value is reached minimum.In order to improve convergence rate, often in network Momentum term is added in weighed value adjusting：

Network weight when middle w (n+1) is (n+1)th iteration, η are learning rate, and E is error energy function, α₀For momentum .

The forecast model obtained using training calculates NOx emission predictive valueBy predicted valueWith actual measured value y^*Carry out Compare, and the fitness value fitness using error sum of squares MSE as particle, fitness calculation formula are as follows：

Initialization extreme value of the fitness value being calculated of each particle as the particle in itself, each population it is central Value minimum MSE is as colony's extreme value.

4.4) more new particle.According to newest individual extreme value and colony's extreme value, according to (4) formula and (5) formula more new particle Speed v_idAnd position x (t)_id(t)：

v_id(t+1)=ω v_id(t)+c₁r₁(p_id-x_id(t))+c₂r₂(p_gd-x_id(t)) (4)

x_id(t+1)=x_id(t)+v_id(t+1) (5)

In formula, t is the optimizing algebraically of particle swarm optimization algorithm, r₁, r₂It is the random number between [0,1], p_idRefer to I-th of particle is sought so far searches optimal location, p_gdRefer to the population and search for optimal location so far.Further, it is The defects of improvement basic particle group algorithm is easily trapped into local extremum and slow convergence rate, introduces on the basis of PSO algorithms Dynamic aceleration pulse c₁、c₂With inertia weight ω：

Wherein, T_maxFor maximum optimizing algebraically, ω_maxFor maximum inertia weight, ω_minFor minimum inertia weight, R₁、R₂、R₃、 R₄For constant.

4.5) particle fitness value calculation.The fitness value of particle after renewal is calculated according to formula (3).

4.6) more new individual extreme value and colony's extreme value.Using fitness value as evaluation index, more contemporary particle and previous generation Fitness value size between particle, if the fitness value of current particle is better than previous generation, the position of current particle is set Individual extreme value is set to, otherwise individual extreme value is kept constant.While the optimal particle of the present age all particle fitness values is obtained, and with Previous generation optimal particles are compared, if the fitness value of contemporary optimal particle is better than the fitness of previous generation optimal particles Value, then be arranged to global optimum by the adaptive optimal control angle value of contemporary particle, and otherwise global optimum is kept constant.

4.7) judge whether to need to carry out simplex search.A simplex search is carried out every 10 generations, is entered if desired Row simplex search then performs step 4.8), otherwise performs step 4.9).

4.8) simplex search.Simplex method (Simplex Method, SM) is a kind of traditional local search of algorithm, it Amount of calculation is small, search speed is fast, has very strong local search ability, can largely make up PSO local optimal searching energy The awkward situation of power relative deficiency.The method constructs a polyhedron with D+1 summit first in D dimension spaces, obtains each summit Adaptive value, and optimum point therein, secondary advantage and most not good enough are determined, then by the strategy such as reflecting, expanding, shrinking or compressing Find out one it is more better, substitution it is most not good enough, so as to form new polyhedron, such iteration can find or approach one most Advantage.Jacobi matrix and Hessian matrix of this method without solved function, without carrying out complicated matrix operation, have Very strong universality.Initial simplex { x is constructed first⁰,x¹,…,xⁱ,…,x^D, x⁰Searched for every sub- population optimal Solution, xⁱAccording to formula (9), (10) generation：

K=-0.05+0.1r (10)

J represents jth dimension variable in formula, and r is obeyed as equally distributed random number on [0,1].

The size of D+1 summit of simplex according to target function is renumberd, meets the numbering on summit：

fitness(x⁰)≤fitness(x¹)≤…≤fitness(xⁱ)≤…≤fitness(x^D) (11)

OrderIfThen stop iteration output x⁰。

4.9) immigrant's operation.In nature, the different population of a species is distributed in different regions, on the one hand each Groupy phase has facilitated their different existence to independently striving for resource from nature used in oneself under different regional conditions Pattern and evolution degree；On the other hand, possibly through migrating between each population, make to contact each other, reach mutual It is connected with nothing, learns from other's strong points to offset one's weaknesses, the purpose of common evolutionary.(Multi-Population, MP) the hereditary grain on multiple populations that this patent uses Swarm optimization is exactly this phenomenon for having used for reference generally existing in nature.Each time in searching process, using single between population Immigrant's operation is carried out to the mode of circulation migration, the excellent individual in the 1st population moves to the 2nd, and the 2nd is moved to the 3rd, with This analogizes, and to the last one is moved to first.Migration rate P between population_i=0.04, i.e., P before ranking in expression source population_i P after ranking among × 100% individual replacement target population_i× 100% individual, optimal knowledge between population is completed with this Exchange.

4.10) algorithm stop condition algorithm judges.Judge whether to reach maximum iteration or reach precision of prediction It is required that the return to step 4.4 if being not reaching to), continued search for using the cluster radius of renewal, otherwise exit search, performed Step 4.11).

4.11) optimal weight threshold particle is exported.

4.12) the initial weight threshold value among optimal particle is assigned to BP neural network model, and combined training sample enters Row study.

4.13) precision of prediction of model is verified.The predicted value of model and actual value are contrasted, relative prediction is calculated and misses Difference.

4.14) relative prediction residual is judged whether within ± 5%, step 2.15) is performed if meeting to require, otherwise Return to step 2.12), the parameter of neural network model, and re -training again.

4.15) output meets desired NOx emission predictive model.

5) model adaptation updates.When the error of NOx discharge and model prediction discharge capacity exceedes ± 5%, immediately more New model.

Claims (2)

1. a kind of real-time estimate system of circulating fluid bed domestic garbage burning emission of NOx of boiler, the system and recirculating fluidized bed The Distributed Control System and production management system of boiler are connected, including data communication interface and host computer, right in host computer NOx emission predictive model is trained and updated, then by the model trained by communication interface be sent to Distributed Control System, Production management system, the host computer include：

Signal acquisition module；The module is used to gather operation of the CFB Domestic refuse incinerators when burning specified house refuse Work condition state parameter and performance variable, and form the training sample matrix X (m × n) of refuse thermal value forecast model input variable, m Number of samples is represented, n represents the number of variable；

Data preprocessing module；Gross error processing and random crror processing are carried out to X (m × n), reject the open country in training sample Value, unusual service condition is excluded, the unusual service condition includes boiler shutdown, banking fire, batcher blocking, and training sample input variable is passed through It is mapped to after normalized in [0,1] section, the training sample X after being standardized^*(m×n)；

Expertise library module；Training sample is constantly updated using the method for rolling time window so that training sample is protected all the time Hold in newest state, rolling time window method refers to since current time, recalls the time chi of L (unit second) length It is very little；

Model modification determination module；The performance of current predictive model is detected, when relative prediction residual exceedes ± 5%, is then judged Model needs to be updated；

Intelligent modeling module；The module is first with the particle cluster algorithm on multiple populations of introducing Operator of Pattern Search to the first of BP neural network Beginning weights and threshold value carry out optimizing, and obtained optimal initial weight threshold value then is assigned into BP neural network model, and as Basis is trained；Comprise the following steps that：

2.1) initialization algorithm parameter；The parameter of BP neural network model and optimizing algorithm is configured, including BP nerve nets Implicit number of plies hl, node in hidden layer hn, training iterations gen1, learning rate η, the hidden layer neuron activation function class of network Type；The maximum optimizing algebraically T of particle cluster algorithm on multiple populations_max, maximum inertia weight ω_max, minimum inertia weight ω_min, speed is more New coefficients R₁、R₂、R₃、R₄, population quantity pop, the number of particles ind of single species；The factor alpha of simplex algorithm, tighten coefficient θ, spreading coefficient γ, constriction coefficient β and search precision ε；

2.2) population is initialized；By the way of real coding, all weight thresholds of BP neural network model are encoded in order Among a particle, and each weight threshold is generated as to a real number between [0,1] at random；

2.3) individual extreme value and colony's extreme value are initialized；The initial weight and threshold value that are included in each particle are assigned to BP nerve nets Network model, and combined training sample carries out learning training, the forecast model obtained using training calculates NOx emission predictive valueWill Predicted valueWith actual measured value y^*It is compared, and the fitness value fitness using error sum of squares MSE as particle, adapt to It is as follows to spend calculation formula：

<mrow> <mi>f</mi> <mi>i</mi> <mi>t</mi> <mi>n</mi> <mi>e</mi> <mi>s</mi> <mi>s</mi> <mo>=</mo> <mi>M</mi> <mi>S</mi> <mi>E</mi> <mo>=</mo> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>m</mi> </munderover> <msup> <mrow> <mo>(</mo> <mover> <mi>y</mi> <mo>^</mo> </mover> <mo>-</mo> <msup> <mi>y</mi> <mo>*</mo> </msup> <mo>)</mo> </mrow> <mn>2</mn> </msup> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>2</mn> <mo>)</mo> </mrow> </mrow>

Initialization extreme value of the fitness value being calculated of each particle as the particle in itself, MSE is most among each population Low value is as colony's extreme value；

2.4) more new particle；According to newest individual extreme value and colony's extreme value, according to (3) formula and the speed of (4) formula more new particle v_idAnd position x (t)_id(t)：

v_id(t+1)=ω v_id(t)+c₁r₁(p_id-x_id(t))+c₂r₂(p_gd-x_id(t))(3)

x_id(t+1)=x_id(t)+v_id(t+1)(4)

In formula, t is the optimizing algebraically of particle swarm optimization algorithm, r₁, r₂It is the random number between [0,1], p_idRefer to i-th Particle is sought so far searches optimal location, p_gdRefer to the population and search for optimal location so far；Further, in order to improve Basic particle group algorithm is easily trapped into the defects of local extremum and slow convergence rate, and dynamic has been introduced on the basis of PSO algorithms Aceleration pulse c₁、c₂With inertia weight ω：

<mrow> <mi>&amp;omega;</mi> <mo>=</mo> <msub> <mi>&amp;omega;</mi> <mrow> <mi>m</mi> <mi>a</mi> <mi>x</mi> </mrow> </msub> <mo>-</mo> <mfrac> <mrow> <msub> <mi>&amp;omega;</mi> <mrow> <mi>m</mi> <mi>a</mi> <mi>x</mi> </mrow> </msub> <mo>-</mo> <msub> <mi>&amp;omega;</mi> <mrow> <mi>m</mi> <mi>i</mi> <mi>n</mi> </mrow> </msub> </mrow> <msub> <mi>T</mi> <mi>max</mi> </msub> </mfrac> <mi>t</mi> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>5</mn> <mo>)</mo> </mrow> </mrow>

<mrow> <msub> <mi>c</mi> <mn>1</mn> </msub> <mo>=</mo> <msub> <mi>R</mi> <mn>1</mn> </msub> <mo>+</mo> <mfrac> <mrow> <msub> <mi>R</mi> <mn>2</mn> </msub> <mo>&amp;times;</mo> <mi>t</mi> </mrow> <msub> <mi>T</mi> <mrow> <mi>m</mi> <mi>a</mi> <mi>x</mi> </mrow> </msub> </mfrac> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>6</mn> <mo>)</mo> </mrow> </mrow>

<mrow> <msub> <mi>c</mi> <mn>2</mn> </msub> <mo>=</mo> <msub> <mi>R</mi> <mn>3</mn> </msub> <mo>-</mo> <mfrac> <mrow> <msub> <mi>R</mi> <mn>4</mn> </msub> <mo>&amp;times;</mo> <mi>t</mi> </mrow> <msub> <mi>T</mi> <mrow> <mi>m</mi> <mi>a</mi> <mi>x</mi> </mrow> </msub> </mfrac> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>7</mn> <mo>)</mo> </mrow> </mrow>

Wherein, T_maxFor maximum optimizing algebraically, ω_maxFor maximum inertia weight, ω_minFor minimum inertia weight, R₁、R₂、R₃、R₄For Constant；

2.5) particle fitness value calculation；The fitness value of particle after renewal is calculated according to formula (2)；

2.6) more new individual extreme value and colony's extreme value；Using fitness value as evaluation index, more contemporary particle and previous generation particles Between fitness value size, if the fitness value of current particle is better than previous generation, the position of current particle is arranged to Individual extreme value, otherwise individual extreme value holding is constant；Obtain the optimal particle of the present age all particle fitness values simultaneously, and with upper one It is compared for optimal particle, if the fitness value of contemporary optimal particle is better than the fitness value of previous generation optimal particles, The adaptive optimal control angle value of contemporary particle is arranged to global optimum, otherwise global optimum is kept constant；

2.7) judge whether to need to carry out simplex search；A simplex search is carried out every 10 generations, it is single if necessary to carry out Pure shape search then performs step 2.8), otherwise performs step 2.9)；

2.8) simplex search；One polyhedron with D+1 summit of construction, obtains each summit first in D dimension spaces Adaptive value, and determine optimum point therein, secondary advantage and most not good enough, then by reflecting, expanding, shrinking or Compression Strategies are found out One is more better, and substitution is most not good enough, so as to form new polyhedron, such iteration can find or approach one it is optimal Point；Specially：Initial simplex { x is constructed first⁰,x¹,…,xⁱ,…,x^D, x⁰The optimal solution searched for every sub- population, xⁱAccording to formula (8), (9) generation：

<mrow> <msup> <mi>x</mi> <mi>i</mi> </msup> <mrow> <mo>(</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mrow> <mo>(</mo> <mn>1</mn> <mo>+</mo> <mi>k</mi> <mo>)</mo> <msup> <mi>x</mi> <mn>0</mn> </msup> <mo>(</mo> <mi>j</mi> <mo>)</mo> <mo>,</mo> <msup> <mi>x</mi> <mn>0</mn> </msup> <mo>(</mo> <mi>j</mi> <mo>)</mo> <mo>&amp;NotEqual;</mo> <mn>0</mn> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msup> <mi>x</mi> <mn>0</mn> </msup> <mrow> <mo>(</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>+</mo> <msup> <mn>10</mn> <mrow> <mo>-</mo> <mn>6</mn> </mrow> </msup> <mo>,</mo> <msup> <mi>x</mi> <mn>0</mn> </msup> <mrow> <mo>(</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>=</mo> <mn>0</mn> </mrow> </mtd> </mtr> </mtable> </mfenced> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>8</mn> <mo>)</mo> </mrow> </mrow>

K=-0.05+0.1r (9)

J represents jth dimension variable in formula, and r is obeyed as equally distributed random number on [0,1]；

The size of D+1 summit of simplex according to target function is renumberd, meets the numbering on summit：

fitness(x⁰)≤fitness(x¹)≤…≤fitness(xⁱ)≤…≤fitness(x^D) (10)

OrderIfThen stop iteration output x⁰；

2.9) immigrant's operation；(Multi-Population, the MP) particle cluster algorithm on multiple populations used；Searching process each time In, carrying out immigrant's operation by the way of one-way circulation migration between population, the excellent individual in the 1st population moves to the 2nd, 2nd is moved to the 3rd, and by that analogy, to the last one is moved to first；Migration rate P between population_i=0.04, that is, represent With P before ranking in source population_iP after ranking among × 100% individual replacement target population_i× 100% individual, is completed with this The exchange of optimal knowledge between population；

2.10) algorithm stop condition judges；Judge whether to reach maximum iteration or reach the requirement of precision of prediction, if It is not reaching to then return to step 2.4), continued search for using the cluster radius of renewal, otherwise exit search, perform step 2.11)；

2.11) optimal weight threshold particle is exported；

2.12) the initial weight threshold value among optimal particle is assigned to BP neural network model, and combined training sample Practise；

2.13) precision of prediction of model is verified；The predicted value of model and actual value are contrasted, calculate relative prediction residual；

2.14) relative prediction residual is judged whether within ± 5%, and step 2.15) is performed if meeting to require, is otherwise returned The parameter of step 2.12), again neural network model, and re -training；

2.15) output meets desired NOx emission predictive model；

Communication module；The module sends the NOx emission predictive model for meeting to require to function Distributed Control System, production management System.

2. a kind of circulating fluid bed domestic garbage burning emission of NOx of boiler Forecasting Methodology, it is characterised in that this method includes following Step：

1) analyze circulating fluid bed domestic garbage burning boiler operation mechanism and NOx formation mechanisms, select rubbish feeding coal, Coal-supplying amount, primary air flow, secondary air flow, flue gas oxygen content, combustion chamber draft, bed temperature, burner hearth freeboard temperature are arranged as NOx Put the input variable of forecast model；

2) training sample is gathered；The historical data that input variable is gathered from database, or collection are spaced according to set time The operational factor under operating mode is specified, forms the training sample matrix X (m × n) of NOx emission predictive mode input variable, m represents sample This number, n represent variable number, while gather corresponding NOx discharge as model output training sample Y (m × 1)；

3) data prediction；Gross error processing and random crror processing are carried out to X (m × n), those are not reflection to forgo The deceptive information of boiler accidental conditions, unusual service condition is excluded, and the unusual service condition includes boiler shutdown, banking fire, given Material machine blocks, in order to avoid the difference of dimension between the parameter of forecast model and order of magnitude shadow bad to caused by model performance Ring, training sample input variable is mapped to after normalized in [0,1] section, the input variable after being standardized Training sample X^*The training sample Y of (m × n) and output variable^*(m×1)；

4) intelligent algorithm integrated moulding；Calculated first with the population on multiple populations for introducing Operator of Pattern Search to the initial of BP neural network Weights and threshold value carry out optimizing, and obtained optimal initial weight threshold value then is assigned into BP neural network model, and as base Plinth is trained；Algorithm steps are as follows：

4.1) initialization algorithm parameter；The parameter of BP neural network model and optimizing algorithm is configured, wrapped in the step Include the implicit number of plies hl of BP neural network, node in hidden layer hn, training iterations gen1, for learning rate η, hidden layer nerve First activation primitive type；The maximum optimizing algebraically T of particle cluster algorithm on multiple populations_max, maximum inertia weight ω_max, minimum inertia power Weight ω_min, speed renewal coefficients R₁、R₂、R₃、R₄, population quantity pop, the number of particles ind of single species；Simplex algorithm Factor alpha, tighten coefficient θ, spreading coefficient γ, constriction coefficient β and search precision ε；

4.2) population is initialized；By the way of real coding, all weight thresholds of BP neural network model are encoded in order Among a particle, and each weight threshold is generated as to a real number between [0,1] at random；

4.3) individual extreme value and colony's extreme value are initialized；The initial weight and threshold value that are included in each particle are assigned to BP nerve nets Network model, the forecast model obtained using training calculate NOx emission predictive valueBy predicted valueWith actual measured value y^*Compared Compared with, and the fitness value fitness using error sum of squares MSE as particle, fitness calculation formula are as follows：

<mrow> <mi>f</mi> <mi>i</mi> <mi>t</mi> <mi>n</mi> <mi>e</mi> <mi>s</mi> <mi>s</mi> <mo>=</mo> <mi>M</mi> <mi>S</mi> <mi>E</mi> <mo>=</mo> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>m</mi> </munderover> <msup> <mrow> <mo>(</mo> <mover> <mi>y</mi> <mo>^</mo> </mover> <mo>-</mo> <msup> <mi>y</mi> <mo>*</mo> </msup> <mo>)</mo> </mrow> <mn>2</mn> </msup> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1</mn> <mo>)</mo> </mrow> </mrow>

Initialization extreme value of the fitness value being calculated of each particle as the particle in itself, the central MSE of each population Minimum value is as colony's extreme value；

4.4) more new particle；According to newest individual extreme value and colony's extreme value, according to (2) formula and the speed of (3) formula more new particle v_idAnd position x (t)_id(t)：

v_id(t+1)=ω v_id(t)+c₁r₁(p_id-x_id(t))+c₂r₂(p_gd-x_id(t))(2)

x_id(t+1)=x_id(t)+v_id(t+1)(3)

In formula, t is the optimizing algebraically of particle swarm optimization algorithm, r₁, r₂It is the random number between [0,1], p_idRefer to i-th Particle is sought so far searches optimal location, p_gdRefer to the population and search for optimal location so far；Further, in order to improve Basic particle group algorithm is easily trapped into the defects of local extremum and slow convergence rate, and dynamic has been introduced on the basis of PSO algorithms Aceleration pulse c₁、c₂With inertia weight ω：

<mrow> <mi>&amp;omega;</mi> <mo>=</mo> <msub> <mi>&amp;omega;</mi> <mrow> <mi>m</mi> <mi>a</mi> <mi>x</mi> </mrow> </msub> <mo>-</mo> <mfrac> <mrow> <msub> <mi>&amp;omega;</mi> <mrow> <mi>m</mi> <mi>a</mi> <mi>x</mi> </mrow> </msub> <mo>-</mo> <msub> <mi>&amp;omega;</mi> <mrow> <mi>m</mi> <mi>i</mi> <mi>n</mi> </mrow> </msub> </mrow> <msub> <mi>T</mi> <mi>max</mi> </msub> </mfrac> <mi>t</mi> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>4</mn> <mo>)</mo> </mrow> </mrow>

<mrow> <msub> <mi>c</mi> <mn>1</mn> </msub> <mo>=</mo> <msub> <mi>R</mi> <mn>1</mn> </msub> <mo>+</mo> <mfrac> <mrow> <msub> <mi>R</mi> <mn>2</mn> </msub> <mo>&amp;times;</mo> <mi>t</mi> </mrow> <msub> <mi>T</mi> <mrow> <mi>m</mi> <mi>a</mi> <mi>x</mi> </mrow> </msub> </mfrac> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>5</mn> <mo>)</mo> </mrow> </mrow>

<mrow> <msub> <mi>c</mi> <mn>2</mn> </msub> <mo>=</mo> <msub> <mi>R</mi> <mn>3</mn> </msub> <mo>-</mo> <mfrac> <mrow> <msub> <mi>R</mi> <mn>4</mn> </msub> <mo>&amp;times;</mo> <mi>t</mi> </mrow> <msub> <mi>T</mi> <mrow> <mi>m</mi> <mi>a</mi> <mi>x</mi> </mrow> </msub> </mfrac> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>6</mn> <mo>)</mo> </mrow> </mrow>

Wherein, T_maxFor maximum optimizing algebraically, ω_maxFor maximum inertia weight, ω_minFor minimum inertia weight, R₁、R₂、R₃、R₄For Constant；

4.5) particle fitness value calculation；The fitness value of particle after renewal is calculated according to formula (1)；

4.6) more new individual extreme value and colony's extreme value；Using fitness value as evaluation index, more contemporary particle and previous generation particles Between fitness value size, if the fitness value of current particle is better than previous generation, the position of current particle is arranged to Individual extreme value, otherwise individual extreme value holding is constant；Obtain the optimal particle of the present age all particle fitness values simultaneously, and with upper one It is compared for optimal particle, if the fitness value of contemporary optimal particle is better than the fitness value of previous generation optimal particles, The adaptive optimal control angle value of contemporary particle is arranged to global optimum, otherwise global optimum is kept constant；

4.7) judge whether to need to carry out simplex search；A simplex search is carried out every 10 generations, it is single if necessary to carry out Pure shape search then performs step 4.8), otherwise performs step 4.9)；

4.8) simplex search；Construct initial simplex { x⁰,x¹,…,xⁱ,…,x^D, x⁰Searched most for every sub- population Excellent solution, xⁱAccording to formula (7), (8) generation：

<mrow> <msup> <mi>x</mi> <mi>i</mi> </msup> <mrow> <mo>(</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mrow> <mo>(</mo> <mn>1</mn> <mo>+</mo> <mi>k</mi> <mo>)</mo> <msup> <mi>x</mi> <mn>0</mn> </msup> <mo>(</mo> <mi>j</mi> <mo>)</mo> <mo>,</mo> <msup> <mi>x</mi> <mn>0</mn> </msup> <mo>(</mo> <mi>j</mi> <mo>)</mo> <mo>&amp;NotEqual;</mo> <mn>0</mn> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msup> <mi>x</mi> <mn>0</mn> </msup> <mrow> <mo>(</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>+</mo> <msup> <mn>10</mn> <mrow> <mo>-</mo> <mn>6</mn> </mrow> </msup> <mo>,</mo> <msup> <mi>x</mi> <mn>0</mn> </msup> <mrow> <mo>(</mo> <mi>j</mi> <mo>)</mo> </mrow> <mo>=</mo> <mn>0</mn> </mrow> </mtd> </mtr> </mtable> </mfenced> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>7</mn> <mo>)</mo> </mrow> </mrow>

K=-0.05+0.1r (8)

J represents jth dimension variable in formula, and r is obeyed as equally distributed random number on [0,1]；

The size of D+1 summit of simplex according to target function is renumberd, meets the numbering on summit：

fitness(x⁰)≤fitness(x¹)≤…≤fitness(xⁱ)≤…≤fitness(x^D) (9)

OrderIfThen stop iteration output x⁰；

4.9) immigrant's operation；The Genetic Particle Swarm Algorithm on multiple populations used；Each time in searching process, using unidirectional between population The mode of circulation migration carries out immigrant's operation, and the excellent individual in the 1st population moves to the 2nd, and the 2nd is moved to the 3rd, with this Analogize, to the last one is moved to first；Migration rate P between population_i=0.04, i.e., P before ranking in expression source population_i× P after ranking among 100% individual replacement target population_i× 100% individual, the friendship of optimal knowledge between population is completed with this Stream；

4.10) algorithm stop condition judges；Judge whether to reach maximum iteration or reach the requirement of precision of prediction, if It is not reaching to then return to step 4.4), continued search for using the cluster radius of renewal, otherwise exit search, perform step 4.11)；

4.11) optimal weight threshold particle is exported；

4.12) the initial weight threshold value among optimal particle is assigned to BP neural network model, and combined training sample Practise；

4.13) precision of prediction of model is verified；The predicted value of model and actual value are contrasted, calculate relative prediction residual；

4.14) relative prediction residual is judged whether within ± 5%, and step 2.15) is performed if meeting to require, is otherwise returned The parameter of step 2.12), again neural network model, and re -training；

4.15) output meets desired NOx emission predictive model；

5) model adaptation updates；When the error of NOx discharge and model prediction discharge capacity exceedes ± 5%, mould is updated immediately Type.