JP4585983B2 - Plant control method and plant control apparatus - Google Patents

Description

  The present invention relates to a boiler plant control system, and more particularly to a boiler control method and apparatus in a power plant.

  In the field of plant control such as boilers, control logic based on PID control has been the mainstream. In addition, many techniques that can flexibly cope with plant characteristics by a supervised learning function represented by a neural network have been proposed. In order to configure a control device using this supervised learning function, it is necessary to prepare a success example as teacher data in advance, and therefore an unsupervised learning method such as a reinforcement learning method has been proposed.

  This reinforcement learning method is a framework of learning control that generates operation signals to the environment so that the measurement signal obtained from the environment becomes desirable through trial and error interaction with the environment such as the controlled object, According to this, even when a successful case cannot be prepared in advance, there is an advantage that a desired behavior can be learned according to the environment by simply defining a desired state.

  In this reinforcement learning method, the evaluation value calculated using the measurement signal obtained from the environment is a clue, and the expected value of the evaluation value obtained from the current state to the future is maximized. Generate an operation signal. Methods for implementing such a learning function include algorithms such as Actor-Critic, Q-learning, and real-time dynamic programming.

  In addition, there is a framework called Dyna-architecture as a reinforcement learning framework developed from the above-described method. This is a method of learning in advance what kind of operation signal should be generated for a model simulating a control target, and determining an operation signal to be applied to the control target using this learning result. However, at this time, it has a model adjustment function for reducing the error between the controlled object and the model.

On the other hand, with the progress of numerical analysis technology, combustion reaction can be reproduced to some extent by calculation, and as a result, it is also possible to build a model using a simulator that simulates a plant (for example, (See Patent Document 1).
JP 2003-281462 A

  As described above, the reinforcement learning method is effective when the characteristics of the controlled object and the control method cannot be formulated in advance, but in plant operation control, the time to collect the results of trial and error operation before model construction In the meantime, there is a problem that the product quality deteriorates and the loss increases. In addition, the trial and error operation at this time may cause an environmental impact due to changes in the properties of the plant waste. However, effective model construction methods for these problems are not considered.

  In Patent Document 1, it is necessary to make the calculation mesh finer in order to increase the calculation accuracy. However, in the case of a large apparatus such as a boiler, the calculation amount becomes enormous and the operating conditions also change continuously. Therefore, the calculation time is long and it is difficult to construct a model in a practical time.

  As described above, the reinforcement learning method is effective for controlling a plant for which control methods cannot be formulated in advance, but it takes a lot of time if data is accumulated by trial operation of the plant for model construction. The desired control performance may not be obtained during the period.

  An object of the present invention is to provide a plant control method and apparatus using a reinforcement learning method with a short model construction period and excellent performance.

In the plant control method having a learning function for learning the relationship between the operation amount and the plant state, and having a function for calculating the operation amount corrected according to the plant state by the learning function, Calculating a relationship between the air flow rate and the CO concentration generated during operation of the plant by a given condition such as a fuel type and a numerical analysis by simulation based on an operation condition such as an air flow rate of the plant; Interpolate discrete calculation results from the step of calculating the relationship between the air flow rate and the CO concentration, and create a continuous model by approximating the change in the CO concentration with respect to the change in the air flow rate into a continuous relationship. And temporarily calculating the continuous model when any data point of the continuous model is removed, and calculating the error from the removed data point Is compared with the threshold value Te, the result of the comparison, when the error exceeds the threshold value, adding data points between the exception data points and to that point before and after, numerical values added data points Reinforcement learning for analyzing and recalculating a continuous model, and reinforcement learning for recalculating a continuous model that has been corrected by including the operation data of the actual plant of the plant in the continuous model recalculated by the reinforcement learning A step of calculating the manipulated value of the air flow rate using the modified continuous model, and a step of outputting the manipulated value of the air flow rate as an operation signal to the plant. This is achieved in such a way that the plant air flow is controlled.

  At this time, an error between the process value calculated by the numerical analysis and the calculated value by the continuous model is evaluated, a new operation condition parameter is created based on the numerical analysis result, and the created new operation condition parameter is used. The numerical analysis may be re-executed to calculate a new process value, and the continuous model may be created again using the old process value and the new process value calculated by the numerical analysis.

  Similarly, at this time, an amount or concentration of at least one kind of fine particles or volatile organic compounds discharged from the plant is used as a process value that defines the state of the plant, and air flow rate, fuel flow rate, At least one of air temperature, air distribution in the burner of the boiler, burner tilt angle of the boiler, or parallel damper angle of the boiler may be used.

Further, the object is to provide a plant control apparatus having a learning function for learning a relationship between an operation amount and a plant state, and a function for calculating an operation amount corrected in accordance with the plant state by the learning function. Means for calculating the relationship between the air flow rate and the CO concentration generated during operation of the plant by given conditions such as shape and fuel type and numerical analysis by simulation based on operating conditions such as the air flow rate of the plant And a continuous model approximated so that the change in the CO concentration with respect to the change in the air flow rate becomes a continuous relationship by interpolating the discrete calculation results by the means for calculating the relationship between the air flow rate and the CO concentration. The means for creating and the continuous model when any data points of the continuous model are removed are temporarily calculated, and the error from the removed data points is calculated. Compared to have value as a result of comparison, when the error exceeds the threshold value, adding data points between the exception data point and to the point before and after, and numerical analysis for additional data points Reinforcement learning for recalculating the continuous model, and reinforcement learning for recalculating the continuous model corrected by including the operation data of the actual plant of the plant in the continuous model recalculated by the reinforcement learning Means for calculating the operation amount of the air flow rate using the modified continuous model, and means for outputting the operation amount of the air flow rate as an operation signal to the plant. This is achieved in such a way that the air flow rate is controlled.

  At this time, means for evaluating an error between the process value calculated by numerical analysis and the calculated value by the continuous model, means for creating a new operation condition parameter based on the error evaluation result, and the created new operation condition parameter A means for re-executing a numerical analysis by using the method to calculate a new process value, and a means for re-creating a continuous model using the old process value calculated by the numerical analysis and the new process value by the continuous modeling means. You may make it prepare.

  Similarly, at this time, an amount or concentration of at least one kind of fine particles or volatile organic compounds discharged from the plant is used as a process value that defines the state of the plant, and air flow rate, fuel flow rate, You may make it use at least 1 sort (s) of air temperature, the air distribution in a boiler burner, the burner tilt angle of a boiler, or the parallel damper angle of a boiler.

  Since the present invention includes the above-described means, the plant can be controlled by the reinforcement learning method using the result of numerical analysis from the time of plant trial operation, and therefore the introduction period of the control device can be shortened.

  Hereinafter, a plant control method and apparatus according to the present invention will be described in detail with reference to embodiments shown in the drawings.

  FIG. 1 shows that the present invention is applied to a boiler facility of a thermal power plant, and a control method for operating the flow rate of air supplied to the boiler of the plant 1 including the boiler facility to minimize the discharged CO concentration is learned. In this case, the plant 1 is provided with an external output interface 2 and an external input interface 3.

  First, the external output interface 2 inputs a signal from the adder 17 and outputs it to the plant 1 to manipulate the air flow rate. You may provide the keyboard and display for a person to operate as needed.

  Next, the external input interface 3 inputs a signal output from the plant 1 and outputs it as operation data of the plant to the operation result database 5, the basic control command calculation means 4, and the operation amount calculation means 15. Also at this time, a keyboard and a display for human operation may be provided as necessary.

  Next, the basic control command calculation means 4 outputs, for example, a basic control command signal for operating the air flow rate of the plant 1, and is configured by a control device configured by a general PID control logic. Yes.

  Further, the operation result database 5 stores plant operation data output from the external input interface 3.

Here, the operation data is a process value that defines the state of the plant 1, and is data representing the relationship between the concentration of CO and NOx with respect to the air flow rate, as shown in FIG. 2, for example. NOx discharged, addition to CO _2, SOx of CO, mercury, fluorine, there is at least one of the amount or concentration of the particulate compounds or volatile organic compounds, such as dust or mist.

  Next, the numerical analysis means 6 is a simulator that simulates the operation of the plant and uses, for example, a simulation method described in Patent Document 1, and is given the shape of the boiler of the plant 1, the type of coal, and the like. Simulation is performed on the basis of the operating conditions such as the air flow rate and the like, and the concentration of CO generated during plant operation is calculated.

  In addition to the air flow described above, there are at least one kind of operating conditions at this time, such as fuel flow, air temperature, air distribution in the boiler burner, boiler burner tilt angle, or boiler parallel damper angle.

  Then, the analysis result by the numerical analysis means 6 is given as, for example, the value of the CO concentration with respect to the air flow rate, and is stored in the numerical analysis database 7. The data stored at this time is in the same format as the operation performance database 5 shown in FIG.

  The continuous modeling means 8 creates a continuous model 9 by approximating the relationship between the discrete air flow rate and the CO concentration stored in the numerical analysis database 7 so as to be a continuous relationship with respect to a change in parameters. . Therefore, the continuous model 9 is created by the continuous modeling means 8 or the continuous model correcting means 12.

  Next, the error evaluation unit 10 evaluates an error between the data stored in the numerical analysis database 7 and the continuous model 9 created by the continuous modeling unit 8. Further, the parameter adding means 11 sets a value of an air flow rate for additionally performing numerical analysis and adds a data point when the error evaluated by the error evaluating means 10 satisfies a certain condition. Then, numerical analysis is performed by the numerical analysis means 6 on the added data point.

  On the other hand, the continuous model correcting means 12 corrects the model of the continuous model 9 using the operation data in the operation result database 5. Then, the control method learning means 13 learns the air flow operation method based on the continuous model 9 by the reinforcement learning method. The learning result is stored in the learning result database 14. An example of the learning result stored in the learning result database 14 at this time is shown in FIG.

  Next, the operation amount calculation means 15 calculates the value of the air flow rate to be operated using the operation data output from the external input interface 3 and the learning result database 14. For example, when the learning result is shown in FIG. 3, if the air flow rate is 0.45, the air flow rate is controlled to be +0.05. The calculated control signal is output to the subtracter 16.

  Therefore, the subtracter 16 receives the output signal of the basic control command calculation means 4 and the output signal of the operation amount calculation processing 15, calculates the difference between these two types of signals, and outputs the difference to the adder 17. At this time, the switch 18 is provided, whereby the output of the subtracter 16 can be disconnected from the input of the adder 17 and the operation can be performed only by the output of the basic control command calculation means 4 as in the conventional case.

  The adder 17 receives the output signal of the basic control command calculation means 4 and the output signal of the subtracter 16 and calculates and outputs the sum of these two types of signals. Therefore, the subtracter 16 and the adder 17 allow the output signal to the plant 1 to be a signal obtained by correcting the output signal of the basic control command calculation means 4 with the output signal of the manipulated variable calculation processing 15.

  Next, the operation of this embodiment will be described with reference to the flowchart of FIG.

  In the embodiment of FIG. 1, in order to control the CO concentration by manipulating the air flow rate of the boiler, it is necessary to know the change in the CO concentration with respect to the change in the air flow rate. Therefore, first, the relationship between the air flow rate and the CO concentration is calculated by numerical analysis (step 101). An example of the calculation result at this time is shown in FIG. The points in the figure are the results of calculating the air flow rate from 0.3 to 0.7 with an interval of 0.1.

  The analysis at this time is preferably performed at as fine an interval as possible, but since it takes a certain amount of time to calculate each point, only discrete data can be actually obtained. Therefore, in order to interpolate a pointless portion, a continuous model approximated so as to have a continuous relationship with respect to a change in parameters is created (step 102).

  The creation of the continuous model at this time includes a method using polynomial approximation and a method using a neural network.

  Here, the characteristic indicated by the broken line in FIG. 5 is the result of creating a continuous model by approximating the data points in this way. According to this, the CO concentration is continuously obtained from the air flow rate given as the continuous model. Can be estimated.

  Next, data to be stored in the learning result database 14 is constructed by the reinforcement learning method (step 103).

  In this embodiment, since the plant is operated using the learning result, the control performance largely depends on the accuracy of the continuous model used for learning.

  Here, in order to increase the accuracy, it is only necessary to increase the number of data points by narrowing the interval of air flow for performing numerical analysis. However, as described above, it is difficult to perform a numerical analysis on a large number of data points in a practical time, so it is important to select data points effectively and suppress the number of data points. For example, in the case of FIG. 5, it is considered that the importance of data is high near the air flow rate 0.5 where the CO concentration is low. Therefore, data points are added by the following procedure.

  FIG. 6 shows a detailed flowchart for creating a continuous model by numerical analysis. Step 201 to step 208 here correspond to step 101 to step 103 in FIG.

  First, in step 201, a point of air flow rate to be calculated by numerical analysis is set. Since the points to be set here are initial settings, the interval between the points and the number of points at this time may be set as appropriate within the change range of the air flow rate in consideration of analysis accuracy, required time, and the like.

  Next, in step 202, the CO concentration at each air flow point is calculated using numerical analysis means. The calculation results are stored in a numerical analysis database.

  Next, in step 203, the discrete data is interpolated to create a continuous model. After that, in step 204, the continuous model when any data point is removed is temporarily calculated, and in step 205, the continuous model is excluded. Calculate the error from the data point.

  FIG. 7 shows an example of the calculation result at this time. In this case, the error between the point of the air flow rate 0.5 and the continuous model created excluding the point of the air flow rate 0.5 is large. It can be seen that a point near 0.5 has a large influence on the continuous model, and it is necessary to calculate at a fine interval in this part.

  Therefore, after step 205, this error is compared with a certain threshold value (step 206), and if the error exceeds this threshold value, it is determined that the data in the vicinity of the point is highly important. Further, the interval is made fine by adding the number of data between the previous and next points (step 207). Then, the CO concentration is calculated for the added data points using a numerical analysis means, and the continuous model is recalculated.

  Thus, data points can be effectively added by repeating steps 202 to 207 until the error becomes small.

  Here, FIG. 8 shows an example of data when a data point is added in step 207 and a continuous model is created again. With the above procedure, accuracy is reduced while suppressing the number of points to be subjected to numerical analysis. It can be understood that a continuous model with high can be created. Then, the process proceeds to step 208 or step 103, and reinforcement learning is performed using the created continuous model.

  However, since the model obtained by the above procedure uses the result of numerical calculation, an error is unavoidable with the actual plant. Therefore, it is desirable to use operation data of the actual plant if possible.

  Therefore, at first, reinforcement learning is performed using a continuous model created by numerical analysis, and when operation data of the actual plant is obtained, the model should be corrected using the operation data.

  Therefore, in this embodiment, the continuous model is corrected by the following procedure, and the operation method is learned again for the corrected model. For this reason, step 104 in FIG. 4 is provided. Then, the continuous model is corrected by the continuous model correcting means. For example, the continuous model is created again by including the operation data of the actual plant in the data created by the numerical analysis. At this time, in order to give priority to the operation data, an appropriate weight may be given in advance.

  FIG. 9 is an example of the case where the operation data is added and the continuous model is corrected in this way, and it can be seen that the tendency of the operation data of the actual plant is reflected as compared with the continuous model of FIG. For example, a continuous model reflecting the tendency of the operation data is provided in the vicinity of the air flow rate at which the operation data is obtained.

  In addition, since a continuous model based on the result of numerical analysis can be created even in a portion without operation data, a model with higher accuracy can be obtained.

  Next, in step 105, reinforcement learning is performed using the corrected continuous model. In re-learning, a method of partially re-learning is known, and according to this method, a control model can be constructed with a small load.

  And the procedure of these steps 104-105 is implemented whenever plant operation data are added. As a result, the continuous model based on the numerical analysis results can be gradually modified to match the characteristics of the actual plant, and the model that controls the air flow rate can also be adjusted to the characteristics of the actual plant. It can be corrected.

  Therefore, the operation amount of the air flow rate is calculated using the result of the reinforcement learning (step 106), and then an operation signal is output to the plant 1 (step 107), and the air flow rate to be supplied to the boiler of the plant 1 is controlled. However, at this time, since it can be controlled based on the analysis result from the plant trial operation by the above procedure, the time to introduction can be shortened, and it can be corrected to the characteristics of the actual plant as the operation data is accumulated. High-performance CO suppression control can be obtained.

  Therefore, according to the above-described embodiment, the model used for learning can be constructed using the results of numerical analysis before plant operation, and therefore the introduction period of plant control can be shortened, and plant operation results can be reduced. The specification performance can be exhibited from the stage where it is not sufficiently accumulated, and the planned suppression of CO can be obtained. In other words, it not only consumes a lot of fuel and raw materials during the trial operation period, but can also deal with the environmental impact of the emissions from the plant.

  In addition, according to the above-described embodiment, the operation condition parameter is created by error evaluation of the numerical analysis result, and the numerical analysis data is added, thereby reducing the model error and performing control with good performance. At this time, it is possible to always perform control with good performance by correcting the model using plant operation data and performing reinforcement learning again.

  Therefore, by applying the invention according to the above embodiment to a boiler of a thermal power plant, it is possible to reduce the risk of increasing the generation of environmentally hazardous substances such as NOx and CO.

  In the present invention, since the plant can be controlled using the result of numerical analysis from the time of plant trial operation, the introduction period of the control device can be shortened. Also, create operating condition parameters by error evaluation of the value analysis results, reduce numerical errors by adding numerical analysis data, modify the model using plant operation data, and perform reinforcement learning again Therefore, it is possible to always perform control with good performance.

  Further, by applying the present invention to a boiler of a thermal power plant, it is possible to reduce the risk of increasing the generation of environmentally hazardous substances such as NOx and CO.

It is a block block diagram which shows one Embodiment of the plant control apparatus by this invention. It is explanatory drawing which shows an example of the driving performance database or numerical analysis database in one Embodiment of this invention. It is explanatory drawing which shows an example of the learning result database in one Embodiment of this invention. It is a flowchart which shows the process sequence by one Embodiment of this invention. It is explanatory drawing of the continuous model in one Embodiment of this invention. It is a flowchart which shows the numerical analysis data addition process procedure by one Embodiment of this invention. It is explanatory drawing of the error evaluation in one Embodiment of this invention. It is explanatory drawing of numerical analysis data addition in one Embodiment of this invention. It is explanatory drawing of continuous model correction in one Embodiment of this invention.

Explanation of symbols

1: Plant (for example, boiler equipment)
2: External output interface 3: External input interface 4: Basic control command calculation means 5: Operation result database 6: Numerical analysis means 7: Numerical analysis database 8: Continuous modeling means 9: Continuous model 10: Error evaluation means 11: Parameter Additional means 12: Continuous model correction means 13: Control method learning means 14: Learning result database 15: Manipulation amount calculation means 16: Subtractor 17: Adder 18: Switch

Claims (4)

In a plant control method having a learning function for learning a relationship between an operation amount and a plant state, and having a function for calculating an operation amount corrected corresponding to the plant state by the learning function,
Calculates the relationship between the air flow rate and the CO concentration generated during operation of the plant based on given conditions such as the shape of the plant and the type of fuel, and numerical analysis by simulation based on operating conditions such as the air flow rate of the plant. And steps to
By interpolating the discrete calculation result calculated in the step of calculating the relationship between the air flow rate and the CO concentration, it is approximated and continuously so that the change in the CO concentration with respect to the change in the air flow rate becomes a continuous relationship. Creating a model,
Temporary calculation of a continuous model when any data point of the continuous model is removed, and an error with the removed data point is calculated and compared with a threshold value. If exceeded , adding data points between the excluded data points and the points before and after that, performing numerical analysis on the added data points, and performing reinforcement learning to recalculate the continuous model;
Reinforcement learning for recalculating a continuous model corrected by including operation data of actual plant equipment in the continuous model recalculated by the reinforcement learning;
Calculating a manipulated variable of the air flow rate using the modified continuous model;
Outputting an operation amount of the air flow rate as an operation signal to the plant,
A plant control method, wherein an air flow rate of the plant is controlled by the operation signal. The plant control method according to claim 1,
The plant is boiler equipment;
The plant control method according to claim 1, wherein at least one of a fuel flow rate of the boiler equipment, an air temperature, an air distribution in a boiler burner, a boiler burner tilt angle, and a boiler parallel damper angle is used as the operation condition. In a plant control device having a learning function for learning the relationship between an operation amount and a plant state, and having a function for calculating an operation amount corrected according to the plant state by the learning function,
Calculates the relationship between the air flow rate and the CO concentration generated during operation of the plant based on given conditions such as the shape of the plant and the type of fuel, and numerical analysis by simulation based on operating conditions such as the air flow rate of the plant. Means to
Interpolating discrete calculation results by means for calculating the relationship between the air flow rate and the CO concentration, to create a continuous model the change in the CO concentration was approximated to be a continuous relationship to the change of the air flow Means,
Temporary calculation of a continuous model when any data point of the continuous model is removed, and an error with the removed data point is calculated and compared with a threshold value. If exceeded, means to add a data point between the excluded data point and the points before and after it, perform numerical analysis on the added data point, and perform reinforcement learning to recalculate the continuous model;
Means for performing reinforcement learning for recalculating a continuous model corrected by including operation data of actual plant equipment in the continuous model recalculated by the reinforcement learning;
Means for calculating the manipulated variable of the air flow rate using the modified continuous model;
Means for outputting an operation amount of the air flow rate as an operation signal to the plant,
A plant control apparatus, wherein an air flow rate of the plant is controlled by the operation signal. In the plant control device according to claim 3,
The plant is boiler equipment;
The plant control apparatus characterized in that at least one of the fuel flow rate of the boiler equipment, the air temperature, the air distribution in the boiler burner, the boiler burner tilt angle, and the boiler parallel damper angle is used as the operation condition.

Images