JP4270218B2 - Control device for control object having combustion device, and control device for plant having boiler - Google Patents

Description

  The present invention relates to a control device for a control object including a combustion device.

  Conventionally, control logic based on PID control has been mainstream in the field of plant control. 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 the control device using the supervised learning function, it is necessary to prepare a success example as teacher data in advance, and thus an unsupervised learning method has been proposed.

  An example of unsupervised learning is reinforcement learning.

  The reinforcement learning method is a learning control framework that generates an operation signal to the environment so that a measurement signal obtained from the environment becomes desirable through a trial and error interaction with the environment such as a control target. Thereby, even when a success example cannot be prepared in advance, there is an advantage that a desired action can be learned according to the environment by simply defining a desired state.

  In reinforcement learning, the expected value of the evaluation value obtained from the current state to the future, based on the evaluation value of the scalar quantity calculated by using measurement signals obtained from the environment (called reward in reinforcement learning) Has a learning function to generate an operation signal to the environment. As a method for implementing such a learning function, there are algorithms such as Actor-Critic, Q-learning, and real-time dynamic programming described in Non-Patent Document 1, for example.

  In addition, a framework called Dyna-architecture is introduced in the above-mentioned document as a framework of reinforcement learning that is an extension of 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 the learning result. It also has a model adjustment function that reduces the error between the controlled object and the model.

  Moreover, in the control apparatus of the plant provided with the combustion apparatus, for example, when the fuel properties are not constant, such as coal fuel, or when the type of coal is changed, there is a problem that the combustion characteristics and heat transfer characteristics of the plant change. is there. As a method for dealing with this problem, for example, a technique described in Patent Document 1 can be cited.

  This is a method for calculating the fuel heating value ratio from the deviation between the measured value and the set value of the main steam pressure in the coal fired boiler.

  Patent Document 2 discloses a first estimation means for calculating an estimated value of heat absorption of a furnace estimated based on fluid measurement data relating to temperature, pressure, flow rate, etc. in a coal-fired boiler furnace, Second estimation means for calculating an estimated heat value of the final recombustor estimated based on fluid measurement data relating to pressure, flow rate, etc., and an estimated value of the absorbed heat quantity of the furnace calculated by the first estimation means, Means for determining the ratio of the heat absorption amount estimated value of the final reburner calculated by the second estimation means, and grasping the combustion characteristics of the boiler based on the ratio of the heat absorption estimated value determined by this means; It is comprised from the calculating means which outputs a gas distribution damper setting value, the rotation speed setting value of a gas recirculation ventilator, and a boiler input acceleration setting value.

JP 2004-190913 A Japanese Patent Laid-Open No. 8-200604 Reinforcement Learning, joint translation by Sadayoshi Mikami and Masaaki Minagawa, Morikita Publishing Co., Ltd., published on December 20, 2000

  The above document estimates and controls the change in fuel heating value by calculating the heat balance from changes in power generation output value, temperature, pressure, etc., and considers the effect on heat transfer performance. However, changes in fuel properties affect not only the heat transfer performance but also the combustion gas composition.

  Increasing NOx, CO, etc. may affect the external environment or increase the load of the exhaust gas treatment device. However, the above literature describes a method that takes into account the effect on the combustion gas composition. Absent.

  Moreover, since the combustion phenomenon is a complex combined phenomenon of the flow of fuel and air (gas), heat transfer and combustion reaction, it is difficult to control its behavior.

  In particular, it has been difficult to derive an appropriate operation method for changes in exhaust gas composition with respect to changes in fuel properties. Even with the above-mentioned reinforcement learning theory, it takes a long learning period to learn the operation method for various fuel properties through trial and error, and the exhaust gas properties deteriorate during the learning period. There was also a possibility.

    An object of this invention is to provide the control apparatus which can control a combustion gas component appropriately also with respect to the change of a fuel composition.

  The present invention relates to basic control command calculation means for inputting measurement data of a control object having a combustion device and calculating an operation command value for the control object, and a plurality of fuel compositions of fuel supplied to the combustion device. Fuel data storage means for storing operation parameters of the combustion device and a data set of the components in the gas, an operation result database for storing past operation result values of the control object, and a past record of the control object. A data creation means for calculating an operation result value and a data distance between the data sets, and determining a data set having the shortest data distance, and using the data set determined by the data creation means, Modeling means for modeling the relationship between the operating parameters and the components in the combustion gas of the combustion device, and using the model of the modeling means, Calculating the operating parameters of the combustion device becomes a good condition, a control unit and a correction means for correcting the operation parameters calculated an operation command value of the basic control command operation unit.

  The present invention comprises the above means, and even if the fuel composition (property) changes, the exhaust gas components can be automatically controlled appropriately, so that the amount of harmful substances such as NOx and CO in the exhaust gas can be reduced.

Hereinafter, the best embodiment will be described with reference to the drawings. FIG. 1 shows a first embodiment. The control device 200 of the present invention receives the process value measurement information 205 from the plant 100 to be controlled, and uses this information to perform a pre-programmed operation in the control device 200 to perform an operation command signal (control) to the plant 100. Signal) 285 is transmitted. The plant 100 controls the state of the plant by operating an actuator such as a valve opening or a damper opening in accordance with the received operation command signal 285.

  This embodiment is an example applied to combustion control of a thermal power plant. In this example, an example applied to a control function aimed at reducing the NOx and CO concentrations in the exhaust gas will be mainly described.

  FIG. 7 shows a configuration of a thermal power plant that is a control target. Coal as fuel, primary air for transporting coal, and secondary air for combustion adjustment are input to the boiler 101 via the burner 102, and the boiler 101 burns the coal. Coal and primary air are led from the pipe 134 and secondary air is led from the pipe 141. Further, after-air for two-stage combustion is introduced into the boiler 101 via the after-air port 103. The after air is guided from the pipe 142.

  Hot gas generated by the combustion of coal flows along the path of the boiler 101 and then passes through the air heater 104. Then, after removing harmful substances with an exhaust gas treatment device, it is discharged from the chimney to the atmosphere.

  The feed water circulating through the boiler 101 is guided to the boiler 101 via the feed water pump 105 and is heated by the gas in the heat exchanger 106 to become high-temperature and high-pressure steam. In the present embodiment, the number of heat exchangers is one, but a plurality of heat exchangers may be arranged.

  The high-temperature and high-pressure steam that has passed through the heat exchanger 106 is guided to the steam turbine 108 via the turbine governor 107. The steam turbine 108 is driven by the energy of the steam, and the generator 109 generates power.

  Next, the paths of the primary air supplied from the burner 102, the secondary air, and the after air supplied from the after air port 103 will be described.

  The primary air is led from the fan 120 to the pipe 130, branched into a pipe 132 that passes through the air heater and a pipe 131 that does not pass through the air, merges again in the pipe 133, and is led to the mill 110. The air passing through the air heater is superheated by the gas. Using this primary air, coal (pulverized coal) generated by the mill 110 is conveyed to the burner 102.

  The secondary air and the after air are led from the fan 121 to the pipe 140 and heated by the air heater 104, and then branched into a secondary air pipe 141 and an after air pipe 142, respectively, and the burner 102 and the after air, respectively. Guided to the airport 103.

  The control device 200 has a function of adjusting the amount of air introduced from the burner and the amount of air introduced from the after air port in order to reduce the NOx and CO concentrations. Although not shown in FIG. 7, there may be a case where a gas recirculation facility for returning a part of the flue gas to the furnace is provided, or a device for changing the burner jet angle up and down. May be the target of the control operation. The fuel flow rate to be supplied to the burner, the burner air flow rate, the air flow rate to be supplied to the air port, the operation amount of the gas recirculation facility, the burner ejection angle, and the like, which are the control operation targets, are the operation parameters for the boiler.

  The control device 200 includes a basic control command calculation unit 230, a correction unit 260 that changes or corrects the basic operation command value 235 output from the basic control command calculation unit 230, a process measurement value 205, an operator input signal, a host control system. Operation result database 240 that stores and stores operation result data including command signals from I / O, an input / output interface 220 for exchanging data with the control target plant 100 or an operator, etc. , And input / output means 221 for inputting a set value, an operation mode, an operation command for manual operation, and the like.

  The basic control command calculation means 230 has a PID (proportional / integral / derivative) controller as a basic component, and is installed in the plant 100 with process measurement values 205, operator input signals, command signals from the host control system, and the like as inputs. The basic operation command value 235 for various operating devices such as valves, dampers, and motors is calculated and output.

  Since the function and configuration of the basic operation command value 235 are the same as those of a conventional thermal power plant control device, description thereof is omitted here.

  A feature of the present invention resides in that a data creation unit 210, a modeling unit 250, a correction unit 260, and a fuel data storage unit 270 are provided. Hereinafter, each function will be described.

  The modeling unit 250 has a function of creating a model that simulates the relationship between the operation parameters such as the fuel flow rate and the air flow rate, and the concentration of a specific component in the exhaust gas in the operation parameters.

  Data 275 is read from the fuel data storage means 270, and an input / output relationship is learned using an error back-propagation method (back propagation method) in a neural network composed of an input layer, an intermediate layer, and an output layer. The configuration and learning method of the neural network are general methods, and these methods may be other methods, and the present invention does not depend on the configuration of the neural network or the learning method. Description is omitted.

  The input data is the air flow rate for each position of the burner and the after-air port, the fuel flow rate for each burner, and the generator output, and the output data is NOx and CO concentration.

  In this example, the relationship between the fuel flow rate, the air flow rate, the power generation output and the NOx and CO concentrations is modeled. However, the present invention does not limit the input items and output items to this. Further, the modeling method is not limited to the neural network, and the model may be created using another statistical method such as a regression model.

  The fuel data storage means 270 stores a plurality of data sets of input data and output data of the modeling means 250 for each type of coal (coal has different properties depending on the place of production).

  The data set includes a past result data extracted from the operation result database 240 and a calculation result calculated in advance by performing a numerical combustion analysis in the boiler.

  A model cannot be created only with operation result data until operation data is accumulated. Of course, a coal type that has no experience in use cannot be modeled. Therefore, in the present invention, combustion numerical analysis is performed with a calculation system that faithfully simulates the design and operating conditions of the target plant, and the result is stored in the fuel data storage means 270.

  Since the combustion phenomenon is a complex combined phenomenon such as the flow of fuel and air (gas), heat transfer and combustion reaction, it is generally difficult to grasp the behavior. However, for example, by conducting basic combustion experiments with varying conditions such as fuel properties (composition, particle size), combustion atmosphere, etc., and modeling the phenomenon from the results, it is large and complex like a boiler in a thermal power plant. Phenomena with complex internal behavior can be analyzed with practical accuracy.

  In addition, there has been a numerical analysis technique in the past, but in order to analyze a large device such as a boiler with a certain degree of accuracy, an enormous number of calculation grids (mesh) is required. It was virtually impossible because it took too long.

  However, for example, by using a numerical analysis technique described in Japanese Patent Application Laid-Open No. 2003-281462, it is possible to perform high-speed analysis while maintaining accuracy, and there has been a recent improvement in computer performance. Numerical analysis of detailed phenomena in large plants can be realized.

  In the combustion numerical analysis, calculation is made for a plurality of coal compositions. Coal composition is determined by conducting composition analysis on typical coal brands (types). Since coal is a natural resource, its composition is often not the same even in the same production area. Therefore, samples of multiple cases are analyzed and the average composition is used.

  In addition, coal is often stored outdoors at power plants, and the moisture content changes depending on the day due to the weather, so the characteristics may change even with the same type of coal.

  Therefore, the combustion numerical analysis is carried out by changing the moisture content for a plurality of cases for the same type of coal, and the results are also stored. Therefore, it is possible to evaluate the degree of influence of the moisture content on the NOx and CO concentrations.

  Next, the data creation unit 210 will be described with reference to FIG.

  In Step 500, a reference value (inter-data distance allowable value) for determining whether or not to change the data set used by the modeling unit 250 is read. The reference value can be input from the keyboard 222, and once entered, the value is stored. Also, the reference value can be changed later.

  In step 510, the results of the air flow rate at each position of the burner and after-airport, the fuel flow rate for each burner, the generator output, the NOx concentration, and the CO concentration from the operation result database 240 for a predetermined period (for example, one month) before to the present time. The value 245 is read.

  In step 520, the numerical analysis result data set used to create the currently used model is read from the fuel data storage unit 270.

  In step 530, the operation result data 245 read in step 510 is read in step 520 so that the air flow rate for each position of the burner and the after-airport, the fuel flow rate for each burner, and the generator output value are the same. Interpolate between the data in the numerical analysis data set and calculate the interpolated values of NOx concentration and CO concentration at that time.

  Although cubic spline interpolation is used as an interpolation method, other interpolation methods may be used.

  Since the numerical analysis data set is discrete data calculated under predetermined conditions, there are many cases where the conditions are not exactly the same as the operation results data. Therefore, such data interpolation is performed to match the conditions with the operation results data. .

  In step 540, the inter-data distance between the data point of the actual value read in step 510 and the interpolated value of the numerical analysis data calculated in step 530 is obtained.

The inter-data distance is defined by the Euclidean distance shown in Equation (1). When the coordinates of two data points P and Q are given by (Xp1, Xp2, Xp3,..., Xpn), (Xq1, Xq2, Xq3,..., Xqn), the square of the distance d _jk between the two points is Desired. Here, Xpi and Xqi as coordinates are the air flow rate at each position of the burner and the after-air port, the fuel flow rate for each burner, the generator output, the NOx concentration, and the CO concentration. Further, j is the fuel data set number, and k is the number of NOx and CO measurement data in the jth fuel data set.

  In step 550, it is determined whether or not the distance to each data point of the operation result data 245 has been calculated for all the fuel composition data sets stored in the fuel data storage means 270.

  If the calculation is completed for all the fuel composition data sets, the process proceeds to step 560. If an uncalculated fuel composition data set remains, the process returns to step 520, the fuel composition data set to be calculated is changed, and the inter-data distance is calculated in the same procedure.

In step 560, for each fuel data set, first, an average distance _{dj_ave} with the operation result data 245 is _obtained by the equation (2).

Next, the data set that minimizes d _{j_ave} is selected.

In step 570, read elaborate reference value in step 500 compares the (allowable value) and d _{J_ave,} if d _{J_ave} is less than the allowable value, the modeling means 250 _d j_ave d j_ave minimal data set number information 215 To exit.

If d _{j_ave} exceeds the allowable value, the process proceeds to step 580.

In step 580, a new data set is created and a model correction command signal is output to the modeling means 250. If _{dj_ave} exceeds the allowable value, it means that the existing fuel composition data set does not match the latest operation result data. Therefore, a new data set is created by adding a data set obtained by adding recent operation record data to the fuel composition data set having the smallest d _{j_ave} . At that time, a data flag is provided so that the numerical analysis data and the operation result data can be distinguished.

  The newly created data set number and model correction command signal information 215 are output to the modeling means 250.

  When receiving the model correction command information 215, the modeling unit 250 refers to the new data set number and recreates the model using the data of the data set. The model creation method is the same as described above, but the model is created by identifying the operation result data by the data flag and placing importance on the weight of the operation result data. Specifically, by increasing the number of times of operation result data input, the model characteristics are reflected more strongly than other numerical analysis data.

  As a result, at least in the vicinity of the data point where the operation record exists, the characteristics close to the operation record data can be modeled, so that the model error is reduced.

The data creation unit 210 is executed at intervals of one week, for example. When _{dj_ave} exceeds the allowable value, the newly accumulated driving performance data for one week is added and the model is created again. Therefore, the ratio of the driving performance data in the data set used for model creation is Increasing and the model characteristics approach the actual driving characteristics.

  Next, the correction unit 260 will be described. The correction unit 260 outputs a simulated operation command signal 265 corresponding to the operation parameter to the modeling unit 250. Modeling means 250 inputs the simulated operation command signal 265 such as the air flow rate at each position of the burner and the after-airport, the fuel flow rate for each burner, and the generator output to the created model, and NOx which is an output value of the model and The CO concentration is calculated and the gas component information 255 is output to the correction means 260.

  In the correction means 260, each operation amount value which is the current plant state is read from the operation data 205, and the air flow rate at each position of the burner and the after-air port is changeable with the time until the next operation on the basis of the state. The simulated operation command signal 265 is output in the range of the width.

  The variable range of the operation amount is obtained from the operation speeds of actuators such as dampers and valves registered in advance and the operation (control) interval. Further, the changeable width of each operation amount is divided into a predetermined number, and the operation amount is changed for all combinations thereof.

The modeling means 250 calculates the NOx and CO concentrations for each simulated operation command signal 265 thus changed. Among them, a simulated operation command signal that minimizes the evaluation value J defined by Equation (3) is extracted. Here, C _NOx and C _CO are calculated values of NOx and CO concentrations, respectively, and A ₁ and A ₂ are coefficients.

The evaluation value calculated by inputting the current measured values of NOx and CO concentration into Equation (3) is J _R. J evaluated by the calculated value of the model is compared with J _R evaluated by the current measurement value, and when the condition of Expression (4) is satisfied, the basic operation command value 235 is corrected and output as the operation command signal 285. .

  When equation (4) holds, it is indicated that the gas component by the model simulation operation command signal 265 is better than the gas component which is the current measurement value.

  A correction method will be described with reference to FIG.

  The subtractor 281 calculates the deviation signal 287 of the simulated operation command signal 265 having the minimum evaluation value J and the basic operation command value 235, and adds it to the basic operation command value 235 by the adder 284 to correct the corrected operation command value. 288 is created.

  If the NOx and CO concentration calculation values 255, which are output values of the modeling means 250, become abnormal due to an abnormality in input data or an arithmetic circuit, the coefficient multiplied by the multiplier 283 is set to zero. As a result, the corrected operation command value 288 becomes equal to the basic operation command value 235, thereby reducing the risk of erroneously outputting an abnormal signal.

  Whether the NOx and CO concentration calculation values 255, which are the output values of the modeling unit 250, are abnormal is determined by checking the upper and lower limits of the input data and output data to the modeling unit 250 and the upper and lower limits of the change rate. If at least one of the upper and lower limit values deviates from the preset upper / lower limit value, the output signal of the switching device 282 is set to 0 to prevent the output of the simulated operation command signal 265 evaluated in a state where there is a possibility of abnormality. The switch 282 sets the output signal to 1 in other cases.

  The switch 286 receives the determination result of Expression (4), selects either the basic operation command value 235 or the correction operation command value 288, and outputs it as the operation command signal 285.

  As described above, by evaluating the relationship between the operation amount and the NOx and CO generation amounts using the model based on the result of the combustion numerical analysis, an operation for reducing the NOx and CO generation amounts is possible.

  Further, even when the fuel property (coal type) changes, an appropriate numerical analysis data set can be selected, so that the accuracy of the model can be maintained high. For this reason, even when the coal type is changed, it is possible to suppress a decrease in control performance with respect to NOx and CO.

  Therefore, conventionally, an operator inputs information on fuel property change into the control device or changes control parameters depending on the experience and knowledge of the operator is automated. Therefore, high-performance and stable operation is possible without depending on the operator's technical capabilities, and the operator's workload can be reduced.

  Furthermore, when the deviation between the numerical analysis result data set prepared in advance and the operation result data is large, a new data set with the operation result data added is created, so the model can be automatically changed to a model close to the operation result data. is there.

In the example of the present embodiment, the control target process value is the NOx and CO concentration. However, the present invention is not limited to this, and includes CO ₂ , SOx, Hg (mercury) amount, fluorine, dust, or mist in the gas. Fine particles and VOC (volatile organic compounds) may be targeted.

  Next, a second embodiment will be described with reference to FIG.

  The difference from the first embodiment described above is that an operation method for reducing NOx and CO is learned using the reinforcement learning means 290.

  The reinforcement learning means 290 has a function of learning an appropriate operation method corresponding to the plant state by the reinforcement learning theory using the operation data accumulated in the operation result database 240.

  A detailed explanation of reinforcement learning theory is described in, for example, “Reinforcement Learning,” translated by Sadayoshi Mikami and Masaaki Minagawa, Morikita Publishing Co., Ltd., published on December 20, 2000. Explain only the concept of learning.

  FIG. 9 shows the concept of control based on reinforcement learning theory. The control device 610 outputs an operation command 630 to the controlled object 600. The controlled object 600 operates according to the control command 630. At this time, the state of the controlled object 600 is changed by the operation according to the control command 630. A reward 620 is received from the controlled object 600 that is an amount that indicates whether the changed state is desirable or undesirable for the controller 610 and how much they are.

  Actually, the information received from the controlled object is the state quantity of the controlled object, and the control device 610 generally calculates the reward based on the information. In general, the reward is set to increase as it approaches a desirable state, and the reward decreases as the state becomes undesirable.

  The control device 610 performs an operation on a trial and error basis, and learns an operation method that maximizes the reward (that is, approaches a desirable state as much as possible), so that an appropriate operation according to the state of the control object 600 ( Control) logic is built automatically.

  The supervised learning theory represented by a neural network needs to provide a success case as teacher data in advance, and is not suitable when a new plant or phenomenon is complicated and a success case cannot be prepared in advance.

  On the other hand, reinforcement learning theory is classified as unsupervised learning, and is applicable to cases where the characteristics of the controlled object are not always clear because it has the ability to generate desirable operations on a trial and error basis. Have advantages.

  In the second embodiment, this reinforcement learning theory is used.

  Reinforcement learning is learned on a trial and error basis, but in the case of plant control, it is difficult to implement trial and error operations directly against the actual plant because of the risk of operation and damage to the manufactured products of the plant. is there. Therefore, in the present invention, an operation characteristic model is created from the operation results of the plant, and this model is used as a learning method.

  The reinforcement learning means 290 outputs a simulated operation command signal 265 composed of the air flow rate at each position of the burner and the after-air port and the fuel flow rate for each burner with respect to the model created by the modeling means 250. The simulated operation command signal 265 corresponds to the operation conditions of the plant, and the upper and lower limit values, the change width (step size), and the maximum change width that can be taken in one operation are set. Each value of the simulated operation command signal 265 is determined at random within a range of possible values.

  The modeling unit 250 inputs the simulated operation command signal 265 to the created model, and calculates the NOx and CO concentrations that become the output data 255.

  The reinforcement learning unit 290 receives the output data 255 of the modeling unit 250 and calculates a reward value.

The reward is defined by equation (5), where R is the reward value, _ONOX is the NOx value, _OCO is the CO value, _SNOx and _SCO are the NOx and CO target set values, k ₁ , k ₂ , k ₃ and k ₄ are positive constants.

As shown in the equation (5), when the NOx and CO values are lower than the target set value, rewards R ₁ and R ₂ are given, and when it is lower than the target set value, the reward is proportional to the deviation. Is supposed to give.

  There are various other methods for defining the reward, and the method is not limited to the formula (5).

The reinforcement learning means 290 learns the combination of the simulated operation command signal 265, that is, the operation amount so that the reward calculated by the equation (5) is maximized, so that NOx,
A combination of manipulated variables for reducing CO can be learned.

  Reinforcement learning means 290 reads driving data 205 at the current time in a state where learning is completed, and outputs an operation amount 295 that maximizes the reward of equation (5) based on the learning result.

  The correction means 260 corrects the basic operation command value 235 and outputs it as an operation command signal 285.

  The correction method is basically the same as in the first embodiment. As shown in the correction circuit in FIG. 13, the difference from the first embodiment shown in FIG. 11 is that the operation amount 295 calculated by the reinforcement learning means 290 is used instead of the simulated operation command signal 265.

  Normally, the output of the switch 282 is set to 1, and the difference between the basic operation command value 235 and the operation amount 295 is added to the basic operation command value 235 to the basic operation command value 235 and the reinforcement learning is performed. The command value is 288.

  The switch 286 normally selects the reinforcement learning command value 288 and outputs it as the operation command value 285.

  However, if any one of the input / output values of the modeling means 250 and the input / output values of the reinforcement learning means 290 deviates from the upper and lower limit values and the change rate limit range, 0 is selected as the output of the switch 282. At the same time, the switch 286 is set to select and output the basic operation command value 235.

  As a result, the abnormal operation command value is prevented from being output twice due to the abnormality of the data or the arithmetic circuit.

  Further, when the control deviation exceeds the predetermined range as a control result by the reinforcement learning command value 288, or when the frequency or duration exceeds the predetermined range, it is determined that the reinforcement learning command value 288 has no effect. Thus, the output of the reinforcement learning command value 288 can be stopped by selecting (processing) the switches 282 and 286 in the same manner as when data is abnormal. Even at this time, the operation can be continued by the basic operation command value 235, and the operation of the plant is not hindered.

  As described above, in the second embodiment, the optimum learning method can be automatically constructed by the reinforcement learning means 290. Further, in the state where learning is completed, when the current driving state data 205 is input, an operation amount that maximizes the reward is output instantaneously, so that the simulated operation command signal 265 is provided at every control timing as in the first embodiment. Therefore, it is not necessary to obtain an appropriate combination of manipulated variables while changing the combination, and the computer load during control can be reduced. As a result, the stability of the computer operation is increased and the reliability of the control device is improved, so that the safety and stability of the plant operation are also increased.

  Next, a third embodiment will be described with reference to FIG.

The difference from the second embodiment is that a state evaluation unit 300 and a data set switching unit 310 are provided. The state evaluation unit 300 monitors a model error which is a deviation between the calculated value 255 of the model created by the modeling unit 250 and the corresponding operation result data 205.

  Processing procedures of the state evaluation unit 300 and the data set switching unit 310 will be described with reference to FIG.

  In step 600, the set value of the allowable value for the model error is read.

  In step 610, the operation result data 205 and the calculated value 255 obtained by inputting the operation amount result value at that time into the model are read.

  In step 620, a deviation (model error) between the operation result data 205 read in step 610 and the model calculation value 255 under the corresponding operation condition is calculated.

  In step 630, the moving average value is calculated for the past time series data of the model error calculated in step 620, and the moving average value and the rate of change of the moving average value in a predetermined time interval are calculated.

  In step 640, the moving average value of the model error calculated in step 630, the change rate value of the moving average value, and the model error for each time are compared with the respective permissible values read in step 600.

  If it is within the allowable range, the process is terminated. If it is outside the allowable range, a data set switching instruction 315 is output to the data creation means 210 and a model change instruction 316 is output to the modeling means 250.

  Step 600 to step 630 are executed by the state evaluation unit 300, and step 640 is executed by the data set switching unit 310.

  FIG. 5 shows a screen output example of the calculation result of the state evaluation means 300.

  A time series graph of model errors is displayed in the graph area 401 on the display screen 400. The graph shows a model error 408 and its moving average value 409 for each time.

  The vertical axis of the graph is the model error (%), and a range value (for example, 0, 100) to be displayed in the input field 403 can be input. The horizontal axis is time, and the date or time is displayed in the display field 404. Further, the display time can be changed on the horizontal axis by operating the scroll bar 402 with a mouse.

  The display period can be selected from a year unit, a month unit, a week unit, a day unit, and a time unit using a display period selection button 405. When any one of the display period selection buttons 405 is selected with a mouse, a display period input window 410 is displayed, and the start time of the period to be displayed can be designated. When the “OK” button is pressed without inputting anything, the display start time is automatically selected according to the display period selected based on the selected current time. When the “return” button is pressed, the input information is canceled.

  With this screen, the transition of the model error can be monitored in time series, and it is easy to grasp the state of the model in use with respect to the allowable value 407 of the model error. To change the setting of the allowable value, a “setting” button 406 is clicked to shift to a setting screen.

  Further, as a result of calculation by the data set switching means 310, when the allowable range is exceeded, a warning is displayed on the screen. The warning contents include “model error allowable value over”, “model error allowable value over duration display”, and “model error change rate allowable value over”.

  When any warning is given, the display screen 400 is automatically displayed to alert the operator. At the same time, the data set switching unit 310 outputs a data set switching instruction 315 to the data creation unit 210 and outputs a model change instruction 316 to the modeling unit 250.

  In particular, in the case of “over tolerance of model error change rate” and “over tolerance of model error”, it can be regarded as a sudden change in plant characteristics. In this case, there is a high possibility that the fuel properties have changed, and the fuel composition data set is changed or a new data set is created by the method described in the first embodiment, and the model is reconstructed. In addition, reinforcement learning is performed on the reconstructed model to automatically follow the state change.

  Accordingly, it is possible to always automatically monitor the tendency of the model error, and it is possible to re-learn the model change and the reinforcement learning based on the monitoring result, so that stable control performance can always be maintained.

  FIG. 6 shows an example in which the data creation unit 210 changes the fuel data set, reconstructs the model, and displays the result of evaluating the model error. Model errors for fuel compositions A to D are displayed. As a reference for selecting a data set by the data creating unit 210, an average value of model errors displayed in FIG. 6 may be calculated, and a fuel composition data set that minimizes the average value may be selected.

  FIG. 10 shows a display screen example 430 of the fuel composition data set selected by the data creation means 210. In the upper part of the screen, the composition of the selected data set is displayed as a pie chart 431 and a table 433.

  Further, on the lower input screen, the operator can input the fuel property actually used. When a plurality of coal types are blended, the coal type name and the blending rate can be input in the input field 435.

  This information can be sent to the analysis center 30. As shown in FIG. 8, the analysis center 30 is connected to a plurality of power plants 50, 51, 52 via a dedicated communication network 40, and can exchange data with each other.

  When the “Send to Analysis Center” button 434 is pressed on the display example 430 of the fuel composition data set, the information input by the operator in the input field 435 is transmitted to the analysis center 30. Further, the data set name selected by the data creation unit 210 and the latest operation result data (operation command value, process value) are simultaneously transmitted to the analysis center 30.

  When the analysis center 30 receives the above information from the power plants 50 to 51, the target plant structure data for the combustion numerical analysis and the received operation amount actual value are input to the analysis model, and fuel composition data, which is one of the calculation conditions. The NOx and CO concentrations are calculated while variously changing the values, and the fuel composition data that minimizes the error from the received actual measured values of NOx and CO are selected.

  The fuel composition data set selected in this way, the model constructed from this data set, and the learning result of reinforcement learning for this model are transmitted to the target power plant via the dedicated communication network 40.

  Upon confirming this received information, the control device 200 stores the newly received fuel composition data set in the fuel data storage means 270, sets the received model in the modeling means 250, and uses the received reinforcement learning result as the reinforcement learning means. Set to 290.

  The model construction method and the reinforcement learning method in the analysis center 30 are the same as those described in the first and second embodiments.

  As described above, when detailed information regarding the fuel change can be obtained, the data set can be updated to a more accurate data set and a model using the data set, so that high-performance control performance can be maintained.

  In this example, after data is received from the power plant, numerical analysis and model construction and reinforcement learning are performed by changing the fuel composition data set at the analysis center 30, but numerical values with various changes in fuel composition are performed in advance. If the analysis is performed in many cases and the results are saved, it is only necessary to select a data set or model from the saved analysis results when the data of fuel property change is received, and a new model can be immediately added to the power plant. It is possible to provide. Therefore, it is possible to shorten the time during which the control performance is lowered due to the fuel change.

  In addition, before receiving fuel property data from the power plant, various fuel properties are analyzed in advance, and a new fuel composition data set, a model built using it, and its model are targeted. It is also possible to transmit the learning results of reinforcement learning to the power plant one by one and store these information at the power plant.

  In this case, the storage capacity, communication load, and communication cost on the power plant side may increase, but there is no need to communicate with the analysis center 30 when changing the coal type, and the model can be quickly changed to a new model or a new reinforcement learning result. Therefore, it is possible to shorten the time when the risk that the control performance deteriorates due to the change of the coal type is generated.

  The following effects can be given as effects of the above embodiment.

  Even if the fuel composition (property) changes, the exhaust gas components can be automatically controlled appropriately, so the amount of harmful substances such as NOx and CO in the exhaust gas can be reduced.

  Moreover, a long learning period is unnecessary, and the effect of the control device can be exhibited from the beginning. In general, there is a possibility that the amount of harmful substance emission may increase in trial and error driving during learning. However, since learning by trial and error driving is not necessary in the present invention, the amount of harmful substance emission can be reduced.

  Furthermore, since the amount of NOx in the exhaust gas is reduced, the amount of utilities such as the amount of ammonia used in the denitration device can be reduced, and the size of the denitration device can be reduced and the life of the catalyst can be extended.

  In addition, since it can automatically follow changes in fuel properties, the operator's adjustment workload can be reduced, and appropriate control can be realized without depending on the experience and knowledge of the operator, improving the reliability of plant operation. There are also benefits.

  In the said Example, although demonstrated as a control apparatus of the plant which has a boiler mainly, this control apparatus can be used when controlling the control target object which has a combustion apparatus.

It is a figure explaining the structure of the control apparatus which concerns on 1st embodiment of this invention. It is a figure explaining the structure of the control apparatus which concerns on 2nd embodiment of this invention. It is a figure explaining the structure of the control apparatus which concerns on 3rd embodiment of this invention. It is a figure explaining the calculation procedure of a data preparation means. It is a figure explaining the example of a display screen of a model error. It is a figure explaining the example of a display screen of a model error. It is a figure explaining the structure of a thermal power plant. It is a figure explaining the communication relation of a power station group and an analysis center. It is a figure explaining the concept of reinforcement learning. It is a figure explaining an example of a fuel data set display and a fuel property input screen. It is a figure explaining a correction circuit. It is a figure explaining the processing procedure of the state evaluation means 300 and the data set switching means 310. It is a figure explaining a correction circuit.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 ... Plant, 200 ... Control apparatus, 210 ... Data preparation means, 220 ... Input / output interface, 221 ... Input / output means, 230 ... Basic control command calculation means, 240 ... Operation performance database, 250 ... Modeling means, 260 ... Correction Means, 270 ... Fuel data storage means.

Claims (8)

Basic control command calculation means for inputting measurement data of a control object having a combustion device and calculating an operation command value for the control object;
Fuel data storage means for storing operation parameters of the combustion device and a data set of components in the gas for a plurality of fuel compositions of fuel supplied to the combustion device;
An operation record database for storing past operation record values of the control object;
A data creation means for calculating a past driving performance value of the control object and a data distance between the data sets, and determining a data set having a shortest data distance;
Modeling means for modeling the relationship between the operation parameters of the combustion device and the components in the combustion gas of the combustion device, using the data set determined by the data creation means;
A correction unit that calculates an operation parameter of the combustion device that is a better condition than the current gas component using the model of the modeling unit, and corrects the operation command value of the basic control command calculation unit with the calculated operation parameter; The control apparatus of the control target object which has a combustion apparatus characterized by having. In the control apparatus of the control target object which has a combustion apparatus of Claim 1,
The correction means corresponds to an operation parameter with respect to the model created by the modeling means when calculating the operation parameter of the combustion apparatus that is a better condition than the current gas component using the model of the modeling means. An operation for calculating a component in the gas by inputting a plurality of simulated operation signals and calculating a reward value for an operation parameter by at least a predetermined method using the calculated value to maximize or minimize the reward value A control device for a controlled object having a combustion device, comprising reinforcement learning means for learning parameter values based on reinforcement learning theory. In the control apparatus of the control target object which has a combustion apparatus of Claim 1,
The calculated value of the component in the gas calculated by the model of the modeling means and the measured value corresponding thereto are calculated to calculate a deviation, a temporal change rate of the deviation amount, a maximum value or a minimum value of the deviation amount, A control device for a controlled object having a combustion device, characterized by comprising state evaluation means for calculating at least one model error information among a time average value of deviation amounts and variances of deviation amounts. In the control apparatus of the control target object which has a combustion apparatus of Claim 3,
A control object having a combustion apparatus, comprising: a data set switching determination unit that determines whether or not a new model is created by the modeling unit based on model error information calculated by the state evaluation unit Control device. In the control apparatus of the control target object which has a combustion apparatus of Claim 1,
The data creation means includes a data point in a multidimensional space consisting of the operation parameter command value or measurement value of the control object and a measurement value of the gas component, and a data point stored in the fuel data storage means or A feature is that a new data set is created by adding measured values for a predetermined period to a fuel composition data set having a minimum sum or average distance between data points and an interpolated point between the data points. The control apparatus of the control object which has a combustion apparatus to perform. In the control apparatus of the control target object which has a combustion apparatus of Claim 1,
Input means for inputting fuel property information including at least one of a fuel composition, a fuel type name or a blending ratio of a plurality of fuel types, and means for transmitting the fuel property information to the outside of the control device using communication means The control apparatus of the control target object which has a combustion apparatus characterized by the above-mentioned. In the control apparatus of the control target object which has a combustion apparatus of Claims 1-6,
For the plurality of fuel compositions, the operating parameters of the plant and the fine particles comprising nitrogen oxides, carbon monoxide, carbon dioxide, sulfur oxides, mercury, fluorine, soot or mist in the combustion gas at the operating conditions, volatilization A means for receiving a numerical analysis result of a gas component composed of at least one of an organic compound from outside the control device;
A control device for a control object having a combustion device, wherein the received numerical analysis result is stored in the fuel data storage means. In the control apparatus of the control target object which has a combustion apparatus of Claim 1,
The control device of the controlled object having the combustion device is a plant having a boiler,
The operation parameter is any one of a fuel flow rate supplied to the boiler burner, an air flow rate supplied to the boiler burner, and an air flow rate supplied to the boiler air port,
The plant control apparatus having a boiler, wherein the gas component is any one of nitrogen oxides, carbon monoxide, carbon dioxide, sulfur oxides, mercury, fluorine, fine particles, and volatile organic compounds.

Images