An industrial boiler combustion optimization method based on deep recurrent neural network and evolutionary computation

Abstract

The present invention discloses an industrial boiler combustion optimization method based on deep recurrent neural networks and evolutionary computing: (1) Offline nonlinear modeling of boiler combustion: a recurrent neural network structure is used to build a boiler combustion nonlinear model BM, and the training parameters are trained on the BM using an offset data structure to obtain a nonlinear model BM_trd after training; Online optimization of boiler combustion decision: an online decision optimization method is established based on the nonlinear model BM_trd and the offset data structure to find the optimal combination of decision changes so that the thermal efficiency at the next moment is optimized. This method realizes online combustion optimization of industrial boilers when there are fewer types of industrial boiler operating parameters, and the optimization decision results are directly or indirectly used for boiler operation adjustment, making the boiler operation cleaner and more efficient.

Description

Technical Field

[0001] The present invention relates to the field of boiler combustion technology, and in particular to an industrial boiler combustion optimization method based on deep recurrent neural network and evolutionary computing. Background Art

[0002] Currently, industrial boilers still account for a significant portion of my country's market. However, unlike power plant boilers, industrial boilers lack the sophisticated automated control systems and rich operational monitoring data types available. With increasingly stringent domestic local standards for industrial boiler pollutant emissions and rising boiler operating costs, industrial boilers currently face difficulties in reducing pollutant emissions and achieving economical operation. Regarding pollutant control, equipping each industrial boiler with expensive pollutant treatment equipment is often impractical. Regarding economical operation, industrial boilers still rely primarily on manual control, often focusing solely on successful startup and stable operation, rather than on economical operation.

[0003] With the rapid development of artificial intelligence, boiler combustion optimization technology based on big data can provide optimized operation control strategies for boilers and learn better control experience from historical operation data, thereby replacing or assisting human decision-making and achieving further economical and clean operation of boilers. At present, there are many large-scale boiler combustion optimization methods based on big data, among which the main method is to establish a static combustion prediction model. For example, based on the historical data of steady-state operating conditions, nonlinear prediction models such as feedforward neural networks and support vector machines are established. This type of method has some major defects:

[0004] 1) The model data requires high operating conditions, and it is necessary to screen data under steady-state operating conditions. However, in actual boiler operation, it is usually difficult to maintain steady-state operation, and the operating conditions are often continuously adjusted with changes in external load demand.

[0005] 2) The static model is only suitable for optimization under steady-state boiler conditions. When the boiler load fluctuates, the static combustion model has a poor prediction of the current boiler operation.

[0006] 3) The static model has high requirements on data types because it only considers the nonlinear characteristics between input variables and output variables, but does not consider the characteristics of the variables themselves in the time dimension.

[0007] Currently, relatively few evolutionary computing methods are used based on dynamic models. Some methods using recurrent neural networks often rely on a wide range of boiler monitoring data types, making the models complex and computationally slow. Furthermore, they are not suitable for situations where industrial boiler data is scarce. For example, Chinese Patent Publication No. CN109670629A discloses a method for predicting the thermal efficiency of coal-fired boilers based on a long-short-term memory neural network; and Chinese Patent Publication No. CN110705881A discloses an online method and system for calculating filtration efficiency based on an artificial neural network.

[0008] Therefore, it is difficult to establish a big data industrial boiler combustion optimization method based on dynamic modeling. The difficulty lies in the design of the data structure and its combination with the optimization algorithm. Summary of the Invention

[0009] The purpose of the present invention is to provide an industrial boiler combustion optimization method based on deep recurrent neural networks and evolutionary computing. When there are fewer types of industrial boiler operating parameters, the online combustion optimization of industrial boilers can be realized, and the optimization decision results can be directly or indirectly used for boiler operation adjustment, making the boiler operation cleaner and more efficient.

[0010] The present invention provides the following technical solutions:

[0011] A method for optimizing industrial boiler combustion based on deep recurrent neural networks and evolutionary computation, comprising:

[0012] (1) Nonlinear offline modeling of boiler combustion: A recurrent neural network structure is used to build a nonlinear model of boiler combustion, and the training parameters are trained using an offset data structure to obtain the nonlinear model BM_trd after training.

[0013] (2) Online optimization of boiler combustion decision: An online decision optimization method is established based on the nonlinear model BM_trd and the offset data structure to find the optimal combination of decision changes so that the thermal efficiency at the next moment is optimized.

[0014] Nonlinear offline boiler combustion modeling requires the use of recurrent neural network structures, such as standard recurrent neural network layers (RNNs), long short-term memory neural network layers (LSTMs), and gated recurrent neural network layers (GRUs). The constructed boiler combustion nonlinear model (BoilerModel, BM) is a stacked neural network layer structure, consisting of a single convolutional neural network layer (Conv), several recurrent neural network layers, and several fully connected neural network layers (Dense).

[0015] The model, based on the specific boiler model, features different numbers of recurrent and fully connected network layers, nodes in each layer, and node activation functions, making it customizable for each boiler. When the network depth is high, batch normalization layers are used between multiple fully connected layers to optimize the loss distribution of the underlying network. Dropout regularization is applied to the input units of the top-level recurrent neural network layer, and recurrent dropout regularization is applied to the recurrent units to mitigate overfitting during neural network training.

[0016] The BM model has a special training data structure: offset data structure. Since the recurrent neural network has the time series memory characteristic, the input data must contain the time dimension, such as {(X t-2 ,X t-1 ,X t ),X=(x1,x2,…,x n )} represents an n-dimensional input X with a time series length of 3. The offset data structure is defined as follows:

[0017]

[0018] The offset data structure has three types of parameters: 1) Decision parameter to be optimized Control: t-n+1 、Control t and Control t+1 ;2) Boiler operating status parameters Status: Status t-n 、Status t-1 and Status t ;3) Optimize target parameters Target: Target t-n Target t-1 and Target t ;

[0019] In the time dimension, the Control parameter in the input X needs to be one moment ahead of the Status parameter and the Target parameter. The so-called "offset" is the forward offset of the Control parameter in the time dimension (offset to the future direction); the output Y is the Target parameter at the future moment, which is consistent with the moment of the Control parameter; the time series length of X is defined as the time interval without the offset parameter (t-(tn)+1=n+1), and a better value needs to be obtained through debugging; based on the offset data structure, the recurrent neural network can "learn" the feature x i The nonlinear relationship between the variable X and the target Y, as well as the nonlinear relationship between the past, present, and future time series distribution of the variable X and the target Y.

[0020] The three types of parameters in the offset data structure are determined by the boiler user to determine the optimization target parameters. The control parameters and status parameters are determined based on the different target parameters. The determination of these parameters is based on the actual boiler control experience:

[0021] 1) When the Target parameters are thermal efficiency and coal consumption, the Status parameters are: boiler load (heat load or electrical load), main steam pressure (or water supply pressure), boiler combustion chamber temperature gradient (temperature measurements from the burner end to the furnace outlet), and exhaust gas temperature (after the last stage heat exchanger). Control parameters generally depend on the boiler type and mainly optimize air volume, fuel amount, and damper opening, generally focusing on primary air volume, furnace outlet oxygen concentration, fuel supply, and secondary damper opening.

[0022] 2) When the Target parameter is the pollutant concentration, the Status parameter and the Control parameter are based on 1), and the Status parameter should be increased by the properties of the fuel entering the furnace. For example, for NO x For pollutants, volatile matter parameters need to be added. For SO2 pollutants, sulfur parameters need to be added. For other pollutants, the user can customize the fuel properties.

[0023] Since the number of determined data features is relatively small, usually 10-15, it meets the current situation of relatively small data types in industrial boilers. At the same time, it can reduce the number of nodes in the neural network model, build a "lightweight" model, and reduce the amount of calculation during subsequent optimization.

[0024] Finally, the boiler nonlinear offline modeling process is as follows:

[0025] 1) Extracting data sets from the boiler's original operating data, including Target parameters, Control parameters, and Status parameters;

[0026] 2) System error position index: The target value range is manually determined. For example, for thermal efficiency, when the thermal efficiency is greater than 1 or less than 0.7, it is considered invalid. The position index of the invalid data is extracted by traversing the target;

[0027] 3) Data standardization or normalization;

[0028] 4) Random error processing: filtering the data, such as Kalman filtering;

[0029] 5) System error processing: Based on the position index in 2), extract the intermediate data segments with an interval greater than 100 between adjacent position indexes. This operation can separate the errors while ensuring data continuity.

[0030] 6) Establish a sequence pool;

[0031] 7) All data segments extracted in 5) are transformed according to the offset data structure, and each sequence (X, Y) is stored in the sequence pool.

[0032] 8) The order of the sequences in the sequence pool is disrupted to facilitate model convergence. The data is divided into a model training set and a test set in proportion, and the training set is imported into the BM model. The weight parameters of the neural network are continuously iterated and updated through back propagation loss (the error between the true value of the training set and the model prediction value). The optimized model is obtained by continuously adjusting the model hyperparameters (the number of neural network nodes, the number of neural network layers, and the time series length of the offset data structure). This step is offline training. After the training is completed, the optimal weight parameters are saved to obtain the trained nonlinear model BM_trd.

[0033] Based on BM_trd, an online decision optimization method is established. Its process depends on the offset data structure. The specific method is defined as follows:

[0034] 1) Extract the current boiler operation data and create an input sequence of length X c (X_current):

[0035] t is the current time;

[0036] 2) Shift the Control parameter forward to obtain the shifted input X co (X_current_offset):

[0037]

[0038] 3) Establish the optimization object ΔC population: ΔC represents the continuous change of the Control parameter, and its specific optimal value is obtained by evolutionary computation iteration. The range of the algorithm search ΔC is defined as [(ΔC) min ,(ΔC) max ], add ΔC to X co Then get X coi , then the ΔC population is transformed into X coi Population:

[0039]

[0040] Among them, Control t +ΔC=Control t+1 ; (ΔC) min and (ΔC) max It is manually defined, and the maximum and minimum search ranges are generally set to be small to avoid excessive boiler control ranges that may cause operational fluctuations and accidents. The optimal individual in the final ΔC population is the optimal decision change output.

[0041] The evolutionary computing methods used in the boiler combustion decision optimization method include swarm intelligence algorithms and evolutionary algorithms. For different problems, the evolutionary computing method requires the establishment of an appropriate fitness function. The fitness function of this method is defined as follows:

[0042]

[0043] st(ΔC) min ≤ΔC≤(ΔC) max

[0044] Wherein, objV represents the fitness value; Norm(X c ) represents X c The same standardization and normalization process used by BM was adopted; X c [-1,-1] represents the target value of the boiler operation at the current time t, then BM_trd(Norm(X c ))-X c [-1,-1] represents ΔTarget. A negative value indicates that adopting the current ΔC strategy will result in a lower Target in the future, and vice versa. a is the exponential expansion coefficient, which mainly normalizes ΔTarget to (0,+∞) to accommodate the selection operators of some evolutionary algorithms (objV>0). The original ΔTarget is usually small and difficult to distinguish. However, after exponential expansion, the gap in ΔTarget is widened, which is beneficial for evolutionary computing to select strategies ΔC with larger ΔTargets, such as those used in "roulette wheel selection." Finally, a of an appropriate size is obtained based on debugging.

[0045] The combustion optimization process based on evolutionary computing is to continuously iteratively evolve the ΔC population, ultimately optimizing the thermal efficiency at the next moment by finding the optimal combination of decision-making changes. This process is consistent with the actual manual control method of the boiler: first, manually observe the current operating data, such as load and steam pressure, and then fine-tune some decision parameters in the hope that a certain operating parameter of the boiler will be optimized in the future.

[0046] In the nonlinear offline modeling of boiler combustion, the present invention specifies the combustion optimization target by the user, such as boiler thermal efficiency, coal consumption, pollutant emission concentration, etc., and mainly adopts a recurrent neural network structure to establish a nonlinear prediction model based on a small number of determined boiler operation monitoring parameters.

[0047] Because the optimization method provided by this invention requires fewer data types for nonlinear offline boiler combustion modeling, it is suitable for industrial boilers with a relatively limited range of operating parameters. Industrial boilers include small biomass boilers, gas boilers, oil boilers, circulating fluidized bed boilers, and others, but do not include small electric boilers. However, since boilers are generally similar, this combustion optimization method is also applicable to large power plant boilers.

[0048] The optimization method proposed in this paper combines practical boiler control experience with the principles of recurrent neural network algorithms to build a unique offset data structure, enabling integration with evolutionary computation to meet the practical optimization needs of boiler production. Based on an offline model, continuous evolutionary computation optimizes the current boiler operating decisions. The computation typically takes less time than the original operating time interval, meeting the requirements of online boiler optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a structural diagram of the industrial boiler combustion optimization method;

[0050] Figure 2 It is a deep neural network layer stacking structure;

[0051] Figure 3 Detailed flow chart of industrial boiler combustion optimization method;

[0052] Figure 4 This is the fitting diagram of the boiler combustion nonlinear model. DETAILED DESCRIPTION

[0053] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings, but the present invention is not limited to these embodiments.

[0054] An industrial boiler combustion optimization method based on deep recurrent neural network and evolutionary computing, such as Figure 1 As shown in the figure, the system consists of two main parts: 1) nonlinear offline modeling of boiler combustion; 2) online optimization of boiler combustion decisions. For different industrial boilers, by collecting historical data and online data, nonlinear offline modeling and online evolutionary calculation are performed to obtain real-time decision optimization results.

[0055] The boiler combustion nonlinear offline modeling adopts a recurrent neural network structure. In this embodiment, a long short-term memory neural network layer (LSTM_net) is used. Figure 2As shown in the figure, the constructed boiler combustion nonlinear model (BoilerModel, BM) is a stacked neural network structure. After debugging, the final model consists of, from top to bottom, one Convnet layer, three LSTMnet layers, and two Densenet layers. The Convnet extraction sequence length is 2, the output feature sequence length is 30, and the number of nodes in the other layers is 40, 40, 30, 30, and 1, respectively. The Convnet layer uses the ReLU function as the activation function, while the Densenet layers all use the Sigmoid function. Due to the large network depth, BatchNormalization is used between the two Densenet layers to optimize the loss distribution of the underlying network. Dropout regularization is used for the top LSTMnet layer, with a dropout ratio of 0.1.

[0056] The BM model training data uses the "offset data structure", which is defined as follows:

[0057]

[0058] The offset data structure has three types of parameters: 1) the decision parameter to be optimized (Control); 2) the boiler operating status parameter (Status); and 3) the optimization target parameter (Target). In the time dimension, the Control parameter in the input X must be one time ahead of the Status and Target parameters. The so-called "offset" is the forward offset of the Control parameter in the time dimension (offset towards the future). The output Y is the Target parameter at a future time, which is consistent with the time at which the Control parameter is located. The time series length of X is defined as the time interval without the offset parameter (t-(tn)+1=n+1, in this case n=2). Based on the offset data structure, the recurrent neural network can "learn" the feature x. i The nonlinear relationship with target Y, as well as the nonlinear relationship between the past, present, and future time series distributions of variable X and target Y. The three types of parameters in the offset data structure are determined by the boiler user to optimize the target parameters. The control parameters and status parameters are determined based on different target selection types. These parameters are determined based on actual boiler control experience.

[0059] The optimization object of this embodiment is a 130t / h small-scale circulating fluidized bed industrial boiler. The optimization target is thermal efficiency. The Control and Status parameters are determined as follows:

[0060] 1) Status parameters: boiler load, main steam pressure, average bed temperature, average furnace outlet temperature, average exhaust gas temperature;

[0061] 2) Control parameters: left primary air volume, right primary air volume, total coal feed rate, and average oxygen concentration at the furnace outlet.

[0062] Since the number of determined data features is only 10, the number of neural network model nodes after debugging is also relatively small. Finally, the total number of parameters to be trained in the BM model is 34,491. The small number of model parameters helps to reduce the time required for subsequent evolutionary computing iterations.

[0063] Based on the established model structure and data structure, the nonlinear offline modeling of boiler combustion is completed. The process is as follows: Figure 3 As shown in (a):

[0064] 1) Extract several historical operation data of the boiler and establish the total data set History(X,Y);

[0065] 2) System error position index: When the thermal efficiency is greater than 1 or less than 0.7, it is considered invalid. The position index of the invalid data is extracted by traversing History(Y);

[0066] 3) Standardize the total data set History(X);

[0067] 4) Random error processing: Perform Kalman filtering on History(X,Y);

[0068] 5) System error processing: Based on the position index in 2), extract the intermediate data segments with an interval greater than 100 between adjacent position indexes. This operation can separate the errors while ensuring data continuity.

[0069] 6) Establish a sequence pool;

[0070] 7) The intermediate data segments extracted in 5) are transformed according to the offset data structure, and each sequence (X, Y) is stored in the sequence pool.

[0071] 8) Shuffle the sequence pool to facilitate model convergence. Divide the sequence data in the pool into training set and test set in proportion. Import the training set into the BM model, and use the RMSProp gradient descent algorithm to continuously iteratively update the weight parameters of the neural network through back propagation loss (the error between the true value of the training set and the model prediction value). The final training fitting result is as follows Figure 4 As shown in the figure, the ideal fitting line and the model fitting line on the test set overlap well, indicating that the trained model can effectively predict boiler thermal efficiency. This step is offline training. After training, the optimal weight parameters are saved to obtain the trained nonlinear model BM_trd.

[0072] Relying on the offset data structure and combining the boiler control experience of actual controllers, an online decision optimization method is established based on the differential evolution algorithm. The process is as follows: Figure 3 As shown in (b):

[0073] 1) Extract the current boiler operation data and create an input sequence of length X c (X_current):

[0074] t is the current time.

[0075] 2) Shift the Control parameter forward to obtain the shifted input X co (X_current_offset):

[0076]

[0077] 3) Establish the optimization object ΔC population: ΔC represents the continuous change of the Control parameter, and its specific optimal value is obtained by differential evolution iteration. The range of the algorithm search ΔC is defined as [(ΔC) min ,(ΔC) max ], where (ΔC) min =[-3000,-3000,-0.5,-0.5], (ΔC) max =[3000,3000,0.5,0.5]. Add ΔC to X co Then get X coi , then the ΔC population is transformed into X coi Population:

[0078]

[0079] Among them, Control t +ΔC=Control t+1 .

[0080] 4) Calculate individual fitness: X coi After inputting the BM_trd model, the predicted value of thermal efficiency at the next moment is obtained, and the adaptability function is defined as:

[0081]

[0082] st(ΔC) min ≤ΔC≤(ΔC) max

[0083] Wherein, objV represents the fitness value; Norm(X c ) represents X c The same standardization process used by BM was used; X c[-1,-1] represents the thermal efficiency value of the boiler at the current moment, then BM_trd(Norm(X c ))-X c [-1,-1] represents ΔTarget. If it is a negative value, it means that the adoption of the current ΔC strategy will lead to a decrease in future thermal efficiency, and vice versa. a is the exponential expansion coefficient, which is 200. It normalizes ΔTarget to (0,+∞) to adapt to some evolutionary algorithm selection operators (objV>0). At the same time, the original ΔTarget is usually small and difficult to distinguish. After exponential expansion, the gap in ΔTarget is widened, which is conducive to evolutionary computing to select the strategy ΔC with a larger ΔTarget.

[0084] 5) Differential Evolution: If the algorithm convergence conditions are not met, differential mutation, recombination, and selection are performed on the population. The optimization process is to continuously iteratively evolve the ΔC population, finding the optimal combination of decision changes, and ultimately optimizing the thermal efficiency at the next moment. This process is consistent with the actual manual control method of the boiler: first, manually observe the current operating data, such as load and steam pressure, and then fine-tune some decision parameters in the hope that a certain operating parameter of the boiler will be optimized in the future.

[0085] 6) The optimal individual of the ΔC population is the optimal decision change of the output.

[0086] Ultimately, the decision optimization results of this example are shown in Table 1. Optimization was performed for 20 different original operating conditions, with an average calculation time of 130 seconds, significantly shorter than the original operating condition interval (10 minutes), meeting the online optimization requirements. The change in thermal efficiency was generally positive, indicating significant potential for improving thermal efficiency through combustion decision optimization. The optimization results of this example demonstrate the feasibility of this industrial boiler combustion optimization method.

[0087] Table 1 Decision optimization results of the embodiment

[0088]

Claims

1. An industrial boiler combustion optimization method based on deep recurrent neural network and evolutionary computing, characterized in that: The method comprises: (1) Nonlinear offline modeling of boiler combustion: A recurrent neural network structure is used to build a nonlinear model of boiler combustion, and the training parameters are trained using an offset data structure to obtain the nonlinear model BM_trd after training. The offset data structure is: Y=Target t+1 ; The offset data structure has three types of parameters: 1) Decision parameter to be optimized Control: t-n+1 、Control t and Control t+1 ;2) Boiler operating status parameters Status: Status t-n 、Status t-1 and Status t ;3) Optimize target parameters Target: Target t-n Target t-1 and Target t ; In the time dimension, the Control parameter in input X must be one time ahead of the Status and Target parameters. The so-called "offset" is the forward offset of the Control parameter in the time dimension, that is, the offset in the future direction. The output Y is the Target parameter at the future time, which is consistent with the time at which the Control parameter is located. The time series length of X is defined as the time interval without the offset parameter, t-(tn)+1=n+1. (2) Online optimization of boiler combustion decision: An online decision optimization method is established based on the nonlinear model BM_trd and the offset data structure to find the optimal combination of decision changes so that the thermal efficiency at the next moment is optimized.

2. The industrial boiler combustion optimization method based on deep recurrent neural network and evolutionary computing according to claim 1 is characterized in that: The recurrent neural network structure is a common recurrent neural network RNN, a long short-term memory neural network layer LSTM or a gated recurrent neural network layer GRU; the constructed boiler combustion nonlinear model BM is a neural network layer stacking structure, which, from top to bottom, consists of a single-layer convolutional neural network layer Conv, several recurrent neural network layers and several fully connected neural network layers Dense.

3. The industrial boiler combustion optimization method based on deep recurrent neural network and evolutionary computing according to claim 1 is characterized in that: The boiler user determines the optimization target Target parameter, and selects the Control parameter and Status parameter of the determined type according to the different Target parameters, specifically: 1) When the Target parameters are thermal efficiency and coal consumption, the Status parameters are: boiler load, main steam pressure or water supply pressure, boiler combustion chamber temperature gradient, and exhaust gas temperature; the Control parameters are: air volume, fuel volume, and damper opening; 2) When the Target parameter is the pollutant concentration, the Status and Control parameters are based on 1), and the Status parameter adds the properties of the fuel entering the furnace.

4. The industrial boiler combustion optimization method based on deep recurrent neural network and evolutionary computing according to claim 3 is characterized in that: The boiler load is the heat load or the electrical load. The boiler combustion chamber temperature gradient is the temperature measurement value from the burner end to the furnace outlet. The exhaust gas temperature is the exhaust gas temperature after the last stage heat exchanger. The control parameters are: primary air volume, oxygen concentration at the furnace outlet, fuel supply, and secondary air door opening. Furnace fuel properties: when targeting NO x When it comes to pollutants, it is the volatile matter parameter; when it comes to SO2 pollutants, it is the sulfur parameter.

5. The industrial boiler combustion optimization method based on deep recurrent neural network and evolutionary computing according to claim 4 is characterized in that: The process of boiler combustion nonlinear offline modeling is as follows: 1) Extracting data sets from the boiler's original operating data, including Target parameters, Control parameters, and Status parameters; 2) System error position index; 3) Data standardization or normalization; 4) Random error processing; 5) System error processing: Based on the position index in step 2), extract the intermediate data segments with an interval of more than 100 between adjacent position indexes to perform system error processing; 6) Establish a sequence pool; 7) All data segments extracted in 5) are transformed according to the offset data structure, and each sequence (X, Y) is stored in the sequence pool; 8) Shuffle the sequence order in the sequence pool; divide the data into a model training set and a test set in proportion, import the training set into the BM model, and iteratively update the weight parameters of the neural network through backpropagation loss; obtain the optimized model by continuously adjusting the model hyperparameters; this step is offline training. After the training is completed, the optimal weight parameters are saved to obtain the trained nonlinear model BM_trd.

6. The industrial boiler combustion optimization method based on deep recurrent neural network and evolutionary computing according to claim 5 is characterized in that: In step 2), the method of systematic error position indexing is to manually determine the target value range. When it comes to thermal efficiency, when the thermal efficiency is greater than 1 or less than 0.7, it is considered invalid, and the position index of invalid data is extracted by traversing the target. In step 4), random error processing is to filter the data. In step 8), the model hyperparameters include the number of neural network nodes, the number of neural network layers, and the time series length of the offset data structure.

7. The industrial boiler combustion optimization method based on deep recurrent neural network and evolutionary computing according to claim 1 is characterized in that: In step (2), the decision online optimization method is: 1) Extract the current boiler operation data and create an input sequence of length X c (X_current): t is the current moment; 2) Shift the Control parameter forward to obtain the shifted input X co (X_current_offset): 3) Establish the optimization object ΔC population: ΔC represents the continuous change of the Control parameter, and its specific optimal value is obtained by evolutionary computation iteration. The range of the algorithm search ΔC is defined as [(ΔC) min ,(ΔC) max ], add ΔC to X co Then get X coi , then the ΔC population is transformed into X coi Population: Among them, Control t +ΔC=Control t+1 ; 4) Calculate individual fitness: X coi After inputting the BM_trd model, the predicted value of thermal efficiency at the next moment is obtained, and the adaptability function is defined as: s.t.(ΔC) min ≤ΔC≤(ΔC) max Wherein, objV represents the fitness value; Norm(X c ) represents X c Adopt standardization and normalization process; X c [-1,-1] represents the target value of the boiler operation at the current time t, then BM_trd(Norm(X c ))-X c [-1,-1] represents ΔTarget. If it is a negative value, it means that the current ΔC strategy will lead to a lower Target in the future, and vice versa. a is the exponential expansion coefficient, which is used to amplify the fitness value and ultimately obtain the appropriate value based on actual debugging. 5) Differential Evolution: If the algorithm convergence conditions are not met, differential mutation, recombination, and selection are performed on the population. The optimization process is to continuously iteratively evolve the ΔC population, finding the optimal combination of decision changes, and ultimately optimizing the thermal efficiency at the next moment. This process is consistent with the actual manual control method of the boiler: first, manually observe the current operating data, then fine-tune some decision parameters, hoping to optimize a certain operating parameter of the boiler in the future; 6) The optimal individual of the ΔC population is the optimal decision change of the output.

Images