JP5155690B2 - Coal-fired boiler gas concentration estimation device and gas concentration estimation method - Google Patents

Description

  The present invention relates to a gas concentration estimation device and a gas concentration estimation method for a coal fired boiler provided in a thermal power plant, and in particular, to determine the concentration of CO and NOx, which are gas components contained in exhaust gas discharged from a coal fired boiler. The present invention relates to a gas fired boiler gas concentration estimating apparatus and a gas concentration estimating method.

  Coal-fired boilers use coal as fuel. In thermal power plants equipped with coal-fired boilers, the concentrations of CO and NOx, which are environmental pollutants contained in the exhaust gas discharged from coal-fired boilers, are specified values. The following must be suppressed.

  The amount of CO and NOx contained in the exhaust gas of a coal-fired boiler is in a contradictory relationship, and if the combustion air (oxygen) is excessive when burning coal with a coal-fired boiler, the amount of NOx produced will increase. Conversely, if the air is insufficient, the amount of CO generated will increase.

  In recent coal-fired boilers, in order to reduce the concentration produced by both CO and NOx and to improve the combustion efficiency of coal-fired boilers, there is a two-stage combustion system in which combustion air is sent stepwise into the coal-fired boiler It has been adopted.

  In this two-stage combustion control, the optimal combustion state of the coal-fired boiler is adjusted by adjusting the amount of combustion air supplied to the coal-fired boiler and selecting the combustion pattern of the burner installed in the coal-fired boiler. Make.

  Further, adjustments for the optimization of the combustion control (control gain adjustment, burner combustion pattern planning, etc.) are performed offline in advance.

  However, the combustion conditions adjusted in advance in this way are optimized for typical operation modes and are only rough operation plans.

  Moreover, since the characteristics of the coal-fired boiler that is a plant change due to aging, what was the optimal combustion condition at the start of the coal-fired boiler operation gradually deviates from the optimal combustion condition over time.

  On the other hand, optimization of the operation of coal-fired boilers (maximizing combustion efficiency while keeping CO and NOx concentrations within an acceptable range) in response to operating conditions such as load demands that change from time to time and aging degradation From the economic point of view.

  In order to realize the optimization of the operation of this coal fired boiler, it is necessary to be able to simulate on-line the changes in the CO and NOx concentrations in the exhaust gas in response to changes in control demand based on the current operating conditions.

  In other words, with respect to the current operating conditions of coal-fired boilers obtained from measurement data, the combustion efficiency of coal-fired boilers and emissions of environmentally hazardous substances are evaluated when the control demand is changed. A function to search for a control point is necessary.

  There are several methods for estimating the CO and NOx concentrations in the exhaust gas discharged from coal-fired boilers. Based on a learning-type algorithm such as a neural network, changes in operating conditions and gas concentrations using actual machine data There are ways to model relationships with trends.

  For example, in JP 2007-26496A, regarding the creation of a continuous model that simulates the characteristics of a plant used for boiler control, a continuous model is created based on boiler process data, and numerically analyzed process data and actual machine A control method for reducing environmentally hazardous substances in exhaust gas by re-creating a continuous model using the operation data and performing reinforcement control using the re-created continuous model to control the boiler is disclosed. . It is also suggested to use a neural network to create a continuous model.

  In the case of such a modeling method, if there is actual machine data, an estimation model of CO concentration and NOx concentration according to the characteristics of the actual machine can be easily created. In other words, using actual machine data during operation even after the start of operation of the coal fired boiler, an estimation model suitable for the state can be created, which is one of the methods often applied.

JP 2007-264796 A

  However, actual machine data, which is a measured value of a coal-fired boiler, includes a transient state when operating conditions such as output are changed. In this case, the correlation between the measured values is a temporary state, and shows a state different from the correlation after the plant has settled over time.

  When modeling the dynamic characteristics of a plant using a neural network suggested in Japanese Patent Application Laid-Open No. 2007-264796, it is effective to learn the neural network using such actual machine data.

  However, when it is desired to learn the static characteristics of the plant, using actual machine data including such a transient state for learning causes a modeling error. Moreover, the measured value of a plant usually includes a measurement error.

  For example, the temperature and the like can be measured with relatively high accuracy, but the flow rate and the like tend to include measurement errors. In addition, sensor aging also causes measurement errors. That is, as described above, in the actual machine data, data with high accuracy and data with low accuracy and large error are mixed.

  When neural network learning is performed using all real machine data in which data with low accuracy is mixed, the constructed neural network model also has low accuracy.

  As a result, when building a model to estimate the CO / NOx concentration in the exhaust gas discharged from a coal-fired boiler that is a plant by learning the trend of actual machine data using actual machine data mixed with low accuracy data There was a problem that the estimation accuracy of the model was lowered.

  An object of the present invention is to estimate an estimation error of a neural network model due to a measurement error included in actual machine data when a change in CO concentration or NOx concentration in exhaust gas is simulated using a neural network in combustion control of a coal fired boiler. An object of the present invention is to provide a gas concentration estimation device and a gas concentration estimation method for a coal fired boiler that can accurately estimate the gas concentration.

The gas concentration estimation device for a coal fired boiler according to the present invention is a gas concentration estimation device for a coal fired boiler that estimates the concentration of gas components contained in exhaust gas discharged from the coal fired boiler using a neural network. The apparatus includes a process database unit that stores process data of a coal-fired boiler, a filtering processing unit that performs filtering processing to extract data suitable for learning of a neural network from the process data stored in the process database unit, and the filtering processing A neural network learning processing unit for performing a neural network learning process based on data suitable for neural network learning extracted by the unit, and an exhaust gas discharged from the coal fired boiler based on the learning process of the neural network learning processing unit CO concentration in or And a neural network estimation processing section for estimating process the Ox concentration, filtering processing in the filtering processing section, to each signal to be used as an input signal of the neural network with respect to the process data used for learning of the neural network learning processing unit On the other hand, with one input signal as a reference, process data whose values of other input signals are within a predetermined threshold are extracted into a group, and the process data belonging to the group is used as a reference input signal and an estimation target. A function obtained by fitting a relationship with a signal value of a certain gas concentration is obtained, and data to be excluded is specified by using a difference between the value of the fitted function and a signal value of a CO concentration or NOx concentration to be estimated. characterized in that it was.
The method for estimating the gas concentration of a coal-fired boiler according to the present invention is a method for estimating the concentration of a gas component discharged from a coal-fired boiler using a neural network. Is stored in the process database section, and the process data stored in the process database section is subjected to a filtering process for extracting data suitable for learning of the neural network by the filtering process section, and the neural network extracted by the filtering process section is processed. Based on the data suitable for learning, the neural network learning processing unit performs the neural network learning processing, and based on the learning processing of the neural network learning processing unit, the neural network estimation processing unit causes CO in exhaust gas discharged from the coal-fired boiler. Dark Alternatively, the NOx concentration is estimated, and the filtering processing in the filtering processing unit is performed by using one input signal for each signal used as the input signal of the neural network with respect to the process data used for learning of the neural network learning processing unit. As a reference, process data whose values of other input signals are within a predetermined threshold are extracted to form a group, and the process data belonging to the group is used as a reference between the input signal used as a reference and the gas concentration signal value to be estimated. A function obtained by fitting the relationship is obtained, and data to be excluded is specified using a difference between the value of the fitted function and the signal value of the CO concentration or NOx concentration to be estimated .

  According to the present invention, when a change in CO concentration or NOx concentration in exhaust gas is simulated using a neural network in combustion control of a coal fired boiler, an estimation error of a neural network model caused by a measurement error included in actual machine data This makes it possible to realize a gas concentration estimation device and a gas concentration estimation method for a coal fired boiler that can accurately estimate the gas concentration.

  The gas concentration estimation apparatus according to the present invention is intended for a coal-fired boiler installed in a thermal power plant, and uses a neural network to estimate the gas concentrations of CO and NOx contained in the exhaust gas discharged from the coal-fired boiler. It is the gas concentration estimation apparatus which performs.

  The CO and NOx substances contained in the exhaust gas discharged from the coal-fired boiler are limited in concentration in the exhaust gas due to environmental regulations.

  An apparatus for estimating the gas concentration of a coal fired boiler according to an embodiment of the present invention estimates CO concentration and NOx concentration in exhaust gas for various operating conditions of a thermal power plant equipped with a coal fired boiler.

  The control system of the thermal power plant equipped with the coal fired boiler which is the subject of the present invention provides the estimated values of the CO concentration and the NOx concentration with respect to various operating conditions (fuel flow rate, air flow rate, etc.) obtained from the gas concentration estimation device. Based on this, the operating conditions of coal-fired boilers that meet the environmental regulations for exhaust gas and maximize boiler efficiency are planned.

  An apparatus and method for estimating the gas concentration of a coal fired boiler as an embodiment of the present invention will be described below with reference to the drawings.

  FIG. 1 is a schematic diagram showing the configuration of a coal-fired boiler gas concentration estimating apparatus according to an embodiment of the present invention.

  Combustion control of the coal-fired boiler 4 that is the plant shown in FIG. 1 is performed by the control system 2. And the gas concentration estimation apparatus 1 of the coal burning boiler 4 shown in FIG. 1 is installed with the gas concentration estimation apparatus 1 for estimating the CO concentration and the NOx concentration in the exhaust gas discharged from the coal burning boiler 4.

  The gas concentration estimation device 1 captures process data as actual machine data of the coal-fired boiler 4 online via the control system 2, and a process database (process DB) 11 that stores the captured process data in time series, A filtering processing unit 12 for performing filtering processing for extracting data suitable for learning of the neural network from the process data stored in time series in the process DB 11 and a filtering processing result storage for storing the filtering processing result in the filtering processing unit 12 Part 13.

  Further, the gas concentration estimation apparatus 1 is configured to detect CO in exhaust gas discharged from the coal-fired boiler 4 based on data suitable for learning of a neural network extracted by filtering processing and stored in the filtering processing result storage unit 13. A neural network learning processing unit 14 that performs a neural network learning process for estimating the concentration and / or NOx concentration, and a learning result storage unit 15 that stores the coupling coefficient obtained by the learning process of the neural network learning processing unit 14. I have.

  Furthermore, the gas concentration estimation device 1 includes a neural network estimation processing unit 16 that estimates the CO concentration and / or NOx concentration in the exhaust gas discharged from the coal fired boiler 4 based on the learning processing of the neural network learning processing unit 14. ing.

  In the embodiment of the present invention, a control system 2 for controlling combustion of the boiler is installed in the coal-fired boiler 4 to be controlled, and the control system 2 operates the coal-fired boiler 4 with respect to the gas concentration estimation device 1. Conditions are set, and the gas concentration estimation apparatus 1 is configured to calculate the estimated values of the CO concentration and the NOx concentration under these operating conditions.

  The gas concentration estimation device 1 uses the process data of the coal fired boiler 4 taken through the control system 2 when estimating the CO concentration and NOx concentration of the coal fired boiler 4.

  The process data of the coal fired boiler 4 taken in from the control system 2 is sent to the process DB 11 installed in the gas concentration estimation apparatus 1 as described above and stored in time series.

  Then, in the gas concentration estimation device 1, the filtering processing unit 12 takes in the process data of the coal fired boiler 4 stored in the process DB 11 in time series, and extracts only the process data suitable for learning of the neural network. Perform processing.

  In the present embodiment, the process of learning the model is performed after omitting data that is a model error from the actual process data of the coal-fired boiler 4 used for model construction. Selection of data that causes the model error is performed as follows.

  There are usually multiple types of data used as model inputs. Of these, first, pay attention to one input signal, and excluding this input signal, a plurality of combinations of data whose values of other input signals are substantially the same are extracted.

  For a plurality of extracted data, a tendency with respect to a change in the input signal of interest is confirmed. At this time, if there is data deviating from the tendency, it is determined that the data becomes an error factor in modeling, and is excluded from the learning data. The above processing is performed for each type of data used for the input signal.

  Next, specific contents of the filtering process in the filtering processor 12 will be described.

  The filtering processing unit 12 acquires an input signal used as an input of the neural network and measurement values corresponding to the estimation target CO concentration and NOx concentration from the process data of the coal fired boiler 4 stored in the process DB 11. To do.

  As an input signal of the neural network, for example, a primary air flow rate for combustion supplied to the coal fired boiler 4 (primary air fan power in the case of non-measurement), a secondary air flow rate, a load of the coal fired boiler, and fuel There is a flow of coal feeders that supply coal.

  Of these, focusing on one signal as the first procedure, the filtering processing unit 12 groups the process data stored in the process DB 11.

  For example, when attention is paid to the signal of the coal feeder flow rate as one signal in the first procedure, data having a difference within a predetermined threshold is defined as the same group for signal values other than the coal feeder flow rate.

  That is, when a signal other than the coal feeder flow rate is used as an input signal, the filtering processor 12 has different values for the coal feeder flow rate, but the primary air flow rate, secondary air flow rate, and boiler load signals are Data indicating almost the same value within a predetermined threshold is collected and grouped.

  Next, for each group grouped above, the filtering processing unit 12 graphs the relationship between the coal feeder flow rate, which is an input signal, and the CO concentration and NOx concentration that are the target of the estimated gas concentration. The concept of this process is shown in FIG.

  FIG. 2 is a graphing process performed by the filtering processing unit 12. For data belonging to the same group, the horizontal axis represents the input signal coal feeder flow rate, and the vertical axis represents the gas concentration to be estimated. The characteristic graph when is plotted is shown.

  Since the values of the signals other than the coal feeder flow rate signal are substantially the same value within the predetermined threshold as described above, the other signals other than the coal feeder flow rate can be assumed to have the same condition. In this case, the characteristic graph shows the dependency of the data on the flow rate of the coal feeder, that is, the degree of the relationship of the influence on the NOx concentration. In the case of FIG. .

  Next, the function fitting is performed by the filtering processing unit 12 on the data plotted in FIG.

  A curved function 100 shown in FIG. 2A shows a fitting function obtained by the fitting process.

  Next, an error between the fitting function 100 and each plotted data is calculated by the filtering process 12.

  And the data in which the error exceeds the predetermined threshold value by the calculation in the filtering process 12 is different from the other data in the dependency on the flow rate of the coal feeder. The neural network learning processing unit 14 excludes the data from the data used in the modeling process for learning the neural network model.

  For example, in the case shown in FIG. 2A, if the error ΔE of the data 101 exceeds a predetermined threshold value with respect to the curved function 100, it is determined that the error ΔE included in the data 101 is large. The data 101 is excluded from data used in the modeling process.

  Next, as in the case shown in FIG. 2B, the fitting process is performed again on the data that has not been excluded by the fitting process in the filtering process 12.

  The example of FIG. 2B shows a situation where the fitting process 102 is newly performed by excluding the data 101 and the fitting process 102 having a curved shape is obtained.

  Next, an error between the new fitting function 102 and each plotted data is calculated by the filtering process 12. If there is data whose error exceeds the predetermined threshold value, the data that causes an error exceeding the predetermined threshold value is excluded, and the fitting process is performed again.

  If there is no data whose error exceeds a predetermined threshold value, the process for this group is completed. Similar processing is performed for another group.

  The processing described above is performed with respect to each signal set as an input signal of the neural network model (in the example described above, the primary air flow rate, the secondary air flow rate, and the boiler load in addition to the coal feeder flow rate).

  By the filtering process in the filtering process 12 described above, it is possible to exclude data having a large error, which becomes a model error factor.

  The filtering processing results in these filtering processing units 12 are stored in the filtering processing result storage unit 13.

  Next, the neural network learning processing unit 14 performs a neural network learning process for estimating the CO concentration and the NOx concentration based on the filtered data stored in the filtering process result storage unit 13.

  A configuration example of the neural network in the neural network learning processing unit 14 is shown in FIG.

  In the configuration example of the neural network that constitutes the neural network learning processing unit 14 shown in FIG. 3, the flow rate of the coal feeder, the primary air flow rate, and the secondary air flow rate that are the process data of the coal-fired boiler 4 are used as the input signal of the neural network. The boiler load is given, and the CO concentration and NOx concentration in the exhaust gas discharged from the coal fired boiler 4 are set as output signals of the neural network.

  A coupling coefficient representing the relationship between the input value and the output value of the neural network is obtained by learning processing in the neural network constituting the neural network learning processing unit 14.

  The coupling coefficient obtained by the learning process using the neural network in the neural network learning processing unit 14 is stored in the learning result storage unit 15.

  The above description is the content of the learning process for constructing the neural network model for estimating the gas concentration (CO concentration and / or NOx concentration) in the exhaust gas discharged from the coal burning boiler 4.

  In the learning processing by the neural network learning processing unit 14, if the measurement data of the coal burning boiler 4 under various operating conditions is stored in the process DB 11, it is not necessary to process in real time.

  It is also possible to prepare the neural network model by performing the above-described learning process in advance before performing the process of estimating the gas concentration (CO concentration and NOx concentration) in the neural network estimation processing unit 16 described later. It is.

  Next, using the neural network model constructed by the learning processing of the neural network learning processing unit 14, the neural network estimation processing unit 16 uses the concentration of gas (CO concentration and / or NOx concentration) contained in the exhaust gas from the coal burning boiler 4. An operation for estimating the above will be described.

  In the control system 2 of the thermal power plant 4 equipped with the coal-fired boiler shown in FIG. 1, the CO concentration and / or the NOx concentration, which is the gas concentration contained in the exhaust gas of the coal-fired boiler 4 for which the environmental regulation value is set, is reduced. Therefore, the combustion control of the coal fired boiler 4 is optimized.

  This optimization process of combustion control corresponds to the control condition among the process values set to the input of the neural network based on the command signal from the control system 2 by the operating condition setting unit 3 installed in the control system 2. Change the process value.

  For example, when analyzing the change tendency of CO concentration or NOx concentration when the primary air flow rate is changed, the operating condition setting unit 3 changes only the value of the primary air flow rate among the input signals of the neural network, As for the values of the input signals (for example, the coal feeder flow rate, the secondary air flow rate, and the boiler load), the current value is set as it is.

  At this time, based on the estimated values of the CO concentration and NOx concentration obtained in the estimation process by the neural network estimation processing unit 16 (in the above example, the change tendency with respect to the primary air flow rate of the input signal is known), the control system 2 is operated with coal. Determine the optimal control method for boiler combustion in a fired boiler.

  As an optimal control method for boiler combustion, for example, it is conceivable to perform combustion control of the boiler by adjusting the amount of combustion air supplied to the boiler or changing the combustion pattern of the boiler.

  About the result of the process by the neural network estimation process part 16 mentioned above, it is comprised so that it can display and confirm on the display apparatus 5 shown in FIG.

  FIG. 4 shows an example of a display example on the display device 5 installed in the present embodiment shown in FIG. 1. In the display example of the display device 5 shown in FIG. 4A, the filtering processing unit 12 performs the filtering process. The excluded data 101 is illustrated.

  Further, in the display example of the display device 5 shown in FIG. 4B, the estimated value for the NOx concentration of the neural network model obtained by the learning process is indicated by a dotted line by the neural network estimation processing unit 16 of the present embodiment. The trend of the estimated value of the NOx concentration over time is also displayed.

  Further, in the display example of FIG. 4B, the measured value of the NOx concentration indicated by the solid line and the trend of the estimated value based on the neural network model indicated by the dotted line are illustrated in a format for comparison.

  In the present embodiment described above, the CO concentration and / or NOx concentration estimation processing in the exhaust gas for optimizing the boiler combustion control by the control system can be performed with high accuracy.

  That is, according to the embodiment of the present invention, when a change in CO concentration or NOx concentration in exhaust gas is simulated using a neural network in the combustion control of a coal-fired boiler, a neural network caused by a measurement error included in actual machine data is used. It is possible to realize a gas concentration estimation device and a gas concentration estimation method for a coal-fired boiler that can estimate a gas concentration with high accuracy by suppressing an estimation error of the net model and enabling high-precision gas concentration estimation.

  The present invention is applicable to a gas concentration estimation apparatus and a gas concentration estimation method for a coal fired boiler for estimating the concentration of CO and NOx, which are gas components contained in exhaust gas discharged from a coal fired boiler.

Schematic which shows the structure of the gas concentration estimation apparatus of the coal fired boiler which is one Example of this invention. Schematic which showed an example of the graphing process in the filtering process part which comprises the gas concentration estimation apparatus of the Example shown in FIG. Schematic which shows an example of a structure of the neural network in the neural network learning process part which comprises the gas concentration estimation apparatus of the Example shown in FIG. Schematic which shows an example of the display in the display apparatus installed in the Example shown in FIG.

Explanation of symbols

  1: gas concentration estimation device, 2: control system, 3: operating condition setting unit, 4: coal-fired boiler, 5: display device, 11: process database, 12: filtering processing unit, 13: filtering processing result storage unit, 14 : Neural network learning processing unit, 15: learning result storage unit, 16: neural network estimation processing unit.

Claims (4)

In a coal fired boiler gas concentration estimation device that estimates the concentration of gas components contained in exhaust gas discharged from a coal fired boiler using a neural network,
The gas concentration estimation apparatus includes a process database unit that stores process data of a coal-fired boiler, a filtering processing unit that performs a filtering process of extracting data suitable for learning of a neural network from the process data stored in the process database unit, A neural network learning processing unit for performing a neural network learning process based on the data suitable for the neural network learning extracted by the filtering processing unit, and the coal-fired boiler based on the learning process of the neural network learning processing unit. the CO concentration or NOx concentration in the exhaust gas to be discharged and a neural network estimation processing section for estimating process from,
The filtering processing in the filtering processing unit is performed by using one input signal as a reference for each signal used as an input signal of the neural network with respect to the process data used for learning of the neural network learning processing unit. Process data whose values are within a predetermined threshold are extracted into a group, and using the process data belonging to the group, a function is obtained by fitting the relationship between the reference input signal and the signal value of the gas concentration to be estimated. An apparatus for estimating the gas concentration of a coal fired boiler, characterized in that the data to be excluded is specified using the difference between the value of the fitted function and the signal value of the CO concentration or NOx concentration to be estimated. In the gas concentration estimation apparatus of the coal fired boiler of Claim 1,
Extracted by the filtering process of the filtering processor
A coal-fired boiler gas concentration estimation apparatus comprising a display device for displaying data belonging to the same group, a function obtained by fitting, and excluded data . In the method for estimating the gas concentration of a coal fired boiler that estimates the concentration of gas components discharged from a coal fired boiler using a neural network,
Coal-fired boiler process data is stored in the process database section, and the process data stored in the process database section is filtered by the filtering section to extract data suitable for neural network learning. Based on the data suitable for learning of the extracted neural network, the neural network learning processing unit performs the neural network learning processing, and based on the learning processing of the neural network learning processing unit, the neural network estimation processing unit Estimate CO concentration or NOx concentration in exhaust gas to be discharged,
The filtering processing in the filtering processing unit is performed by using values of other input signals based on one input signal for each signal used as a neural network input signal for the process data used for learning in the neural network learning processing unit. Is obtained as a group by extracting process data within a predetermined threshold, and using the process data belonging to the group, a function fitting the relationship between the reference input signal and the signal value of the gas concentration to be estimated is obtained. A method for estimating a gas concentration of a coal fired boiler, wherein data to be excluded is specified using a difference between a value of a fitted function and a signal value of a CO concentration or NOx concentration to be estimated . In the gas concentration estimation method of the coal fired boiler of Claim 3,
A method for estimating the gas concentration of a coal-fired boiler, comprising displaying data belonging to the same group, a function obtained by fitting, and excluded data on a display device .

Images