Thermal power combustion optimization control method and system based on deep learning

Abstract

The invention provides a thermal power combustion optimization control method and system based on deep learning, and relates to the technical field of thermal power combustion optimization control, and the method comprises the steps: recognizing an abnormal scene based on a maximum average difference criterion and a multi-kernel density estimation method, and carrying out the sampling of operation data according to the abnormal scene. A mutual information maximization criterion is adopted to carry out dynamic quantization coding on a sample, and a recurrent neural network is utilized to extract a working condition feature vector. And evaluating a system state according to the working condition feature vector, screening control actions meeting constraint conditions through an adversarial learning network, and calculating strategy gradient optimization to obtain an optimal control strategy. And finally, decomposing the optimal strategy into sub-strategies, constructing a distributed consistency coordination model based on a fuzzy decision tree, determining an optimal execution time sequence, and correcting a control signal through a self-adaptive dead-zone compensator to realize optimal control of thermal power combustion. The thermal power generating unit operation stability and economy can be effectively improved, and pollutant emission is reduced.

Description

Technical Field

[0001] The present invention relates to the technology of optimizing the combustion control of thermal power generation, and particularly to a method and system for optimizing the combustion control of thermal power generation based on deep learning. Background Art

[0002] As a major energy production facility, the combustion efficiency and emission level of thermal power units are directly related to energy utilization efficiency and environmental protection. However, due to the non-linear, strongly coupled, and multi-variable characteristics of the boiler combustion process, traditional experience-based control methods are difficult to achieve optimal control under complex working conditions, resulting in a decline in operating efficiency and an increase in pollutant emissions. Therefore, it is of great significance to study control technologies that can adapt to changing working conditions in real time and precisely optimize the combustion process.

[0003] In recent years, with the rapid development of artificial intelligence technology, deep learning has shown great potential in the field of industrial process modeling and optimization. By combining advanced algorithms such as recurrent neural networks and adversarial learning networks, it is possible to achieve feature extraction, working condition evaluation, and generation of optimal control strategies for the operating data of thermal power units. However, current methods usually fail to fully consider practical problems such as distribution drift of the operating data of thermal power units and identification of abnormal working conditions, which limits the applicability and robustness of the models.

[0004] Therefore, the present invention proposes a method for optimizing the combustion control of thermal power generation based on deep learning to achieve an improvement in the combustion efficiency of thermal power units and effective control of emissions. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for optimizing the combustion control of thermal power generation based on deep learning, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiments of the present invention,

[0007] A method for optimizing the combustion control of thermal power generation based on deep learning is provided, including:

[0008] Detecting the degree of distribution drift of the operating data of the thermal power unit by using the maximum mean discrepancy criterion, identifying the type of abnormal scenario based on the multi-kernel density estimation method, calculating the data sampling weight according to the degree of distribution drift and the type of abnormal scenario, sampling the operating data of the thermal power unit according to the data sampling weight to obtain real-time operating samples, calculating the dynamic quantization threshold of the real-time operating samples based on the mutual information maximization criterion, encoding the real-time operating samples by using the dynamic quantization threshold to obtain a feature encoding sequence, and inputting the feature encoding sequence into a recurrent neural network to extract time series features to obtain a working condition feature vector;

[0009] Calculate the system state evaluation value according to the working condition feature vector, generate a control action candidate set based on the system state evaluation value, use an adversarial learning network to determine whether each control action in the control action candidate set meets the system operation constraint conditions, filter to obtain a set of effective control actions that meet the constraint conditions, calculate the policy gradient of each control action in the set of effective control actions, and optimize the set of effective control actions based on the policy gradient to obtain an optimal control strategy;

[0010] Decompose the optimal control strategy into multiple control sub-strategies, construct an execution cost matrix for the control sub-strategies, calculate the conflict degree of the control sub-strategies using a fuzzy decision tree based on the execution cost matrix, establish a distributed consistency coordination model with the conflict degree as the node weight, obtain the optimal execution timing by iteratively updating the node states, generate control signals for each actuator according to the optimal execution timing, use an adaptive dead zone compensator to compensate the control signals for each actuator, calculate the compensation gain based on the actuator dead zone characteristics and correct the control signals, and input the corrected control signals into the thermal power unit execution terminal to achieve combustion optimization control.

[0011] In an alternative embodiment,

[0012] Use the maximum mean discrepancy criterion to detect the distribution drift degree of the thermal power unit operation data, identify the abnormal scenario type based on the multi-kernel density estimation method, calculate the data sampling weight according to the distribution drift degree and the abnormal scenario type, and sample the thermal power unit operation data according to the data sampling weight to obtain real-time operation samples including:

[0013] Divide the thermal power unit operation data into a historical sample set and a current sample set according to a time window, use the feature mapping function in the reproducing kernel Hilbert space to map the historical sample set and the current sample set to obtain historical mapping features and current mapping features, construct Gaussian kernel functions to perform kernel transformation on the historical mapping features and the current mapping features respectively to obtain historical kernel transformation features and current kernel transformation features, use an unbiased estimation method to calculate the maximum mean discrepancy distance between the historical kernel transformation features and the current kernel transformation features, and use the exponential smoothing method to filter the maximum mean discrepancy distance to obtain the distribution drift degree;

[0014] Construct an initial probability density function of the operating condition feature vector using the kernel density estimation method, determine the benchmark bandwidth of the initial probability density function based on the grid search and cross-validation methods, take the probability distribution of the historical sample set as the prior distribution, construct a trial distribution of the current sample set using the kernel density estimation method, calculate the local data density ratio according to the prior distribution and the trial distribution, update the bandwidth of the initial probability density function based on the local data density ratio to obtain an adaptive bandwidth, substitute the adaptive bandwidth into the initial probability density function to obtain an updated probability density function, perform a logarithmic transformation on the updated probability density function to obtain an anomaly metric value of the operating condition feature vector, and determine the anomaly scenario type based on the anomaly metric value;

[0015] Construct a weight calculation model based on the radial basis function, input the distribution drift degree, anomaly metric value, and anomaly scenario type into the weight calculation model together, optimize the network parameters of the weight calculation model using the backpropagation algorithm to obtain sampling weights, construct a probability distribution function of importance sampling based on the sampling weights, and use the Markov chain Monte Carlo method to extract the operating data of the thermal power unit from the probability distribution function to obtain real-time operating samples.

[0016] In an alternative embodiment,

[0017] Calculate the dynamic quantization threshold of the real-time operating samples based on the mutual information maximization criterion, encode the real-time operating samples using the dynamic quantization threshold to obtain a feature coding sequence, and input the feature coding sequence into a recurrent neural network to extract temporal features to obtain an operating condition feature vector, including:

[0018] Construct a mutual information function between the real-time operating samples and their quantization values, where the mutual information function includes the entropy of the real-time operating samples and the conditional entropy under the given quantization values, perform iterative optimization on the mutual information function using the gradient ascent method, update the quantization threshold based on the learning rate and the mutual information gradient in each iteration, and obtain the optimal quantization threshold when the change value of the mutual information function is less than the preset mutual information function convergence threshold, and use the optimal quantization threshold as the dynamic quantization threshold;

[0019] Set the coding bits and quantization levels based on the dynamic quantization threshold, the number of coding bits is adaptively determined according to the distribution characteristics of the operating data of the thermal power unit, the quantization levels are calculated using the minimum description length criterion, and input the coding bits and the quantization levels into the coding function to map the real-time operating samples to obtain a feature coding sequence;

[0020] Divide the feature coding sequence into multiple subsequences according to the temporal complexity, input each subsequence into a long short-term memory network respectively, obtain the cell state through the collaborative calculation of the forget gate and the input gate, process the cell state through the output gate to obtain the hidden state, and merge the results after processing each subsequence to obtain the cell state sequence and the hidden state sequence;

[0021] Calculate the weights for the hidden state sequence and the cell state sequence respectively using the temporal attention mechanism and the feature attention mechanism, and perform weighted fusion on the state sequences based on the weights to obtain the working condition feature vector;

[0022] Calculate the reconstruction error term, sparse constraint term, and temporal consistency constraint term of the working condition feature vector, combine them to obtain the loss function, and optimize the loss function using the backpropagation algorithm with an adaptive learning rate. When the value of the loss function is less than the preset threshold, obtain the optimized working condition feature vector.

[0023] In an alternative embodiment,

[0024] Calculate the system state evaluation value according to the working condition feature vector, generate a control action candidate set based on the system state evaluation value, and use an adversarial learning network to determine whether each control action in the control action candidate set satisfies the system operation constraint conditions, and screen to obtain an effective control action set that meets the constraint conditions, including:

[0025] Construct a state evaluation model with a two-stream network structure. The two-stream network includes a feature extraction stream and a state mapping stream. Input the working condition feature vector into the feature extraction stream for non-linear transformation to obtain a feature representation, input the feature representation into the state mapping stream to obtain the system state, construct a double value function network, input the system state into the double value function network, and take the minimum value of the outputs of the two networks in the double value function network to obtain the system state evaluation value;

[0026] Construct a hybrid policy network based on the system state and the system state evaluation value, calculate the hybrid weight through the hybrid policy network, generate the mean vector and covariance matrix of each component based on the hybrid weight and the system state evaluation value, and perform a transformation on the standard normal distribution random sampling result with the mean vector and covariance matrix to obtain the control action candidate set;

[0027] Use the expert demonstration actions that meet the system constraints as real samples and the non-constraint actions obtained by random sampling as generated samples, and input them into an adversarial discriminator with a residual network structure. The residual network structure includes an input layer and multiple residual blocks. Obtain the input layer features through the input layer. Each residual block performs two-layer transformation on the input features to obtain intermediate features and add them to the input features to obtain residual features. Stack the input layer features and the residual features and pass them through the discriminator output layer to obtain the real sample discrimination value and the generated sample discrimination value;

[0028] Calculate the discriminator loss function value based on the real sample discriminant value, the generated sample discriminant value, and the gradient penalty term, and update the network parameters of the adversarial discriminator in an alternating optimization manner. Repeat the iteration until the preset number of iterations is satisfied to obtain a trained adversarial discriminator;

[0029] Combine the control action candidate set with the system state and input it into the trained adversarial discriminator to obtain candidate action evaluation scores. Map the candidate action evaluation scores to obtain a discrimination threshold, and select the actions with evaluation scores greater than the discrimination threshold as the effective control action set that meets the system constraint conditions.

[0030] In an alternative embodiment,

[0031] Calculating the policy gradient of each control action in the effective control action set and optimizing the effective control action set based on the policy gradient to obtain an optimal control strategy includes:

[0032] Construct a feature extraction network, input the working condition feature vector into the feature extraction network to obtain a feature representation, input the feature representation into a pre-trained policy optimization model, calculate the execution probability distribution of each control action in the effective control action set, and calculate the state value estimation of each control action;

[0033] During the actual operation of the system, execute each control action and record the corresponding actual reward value. Calculate the temporal difference error between the actual reward value and the state value estimation, and use the temporal difference error as the advantage function;

[0034] Calculate the policy gradient based on the advantage function, and use the policy gradient to construct a Fisher information matrix. Analyze the curvature characteristics of the policy gradient in different directions through the Fisher information matrix, determine the parameter update step size in each direction according to the curvature characteristics, so that the step size in the direction with larger curvature is smaller and the step size in the direction with smaller curvature is larger. Multiply the step size in each direction by the corresponding policy gradient component and accumulate to obtain the parameter update amount;

[0035] Update the parameters of the policy optimization model using the parameter update amount to obtain an optimized policy optimization model. Input the effective control action set into the optimized policy optimization model for evaluation, and select the control action with the highest evaluation value from the effective control action set as the optimal control strategy.

[0036] In an alternative embodiment,

[0037] Decompose the optimal control strategy into multiple control sub-strategies, construct an execution cost matrix for the control sub-strategies, calculate the conflict degree of the control sub-strategies using a fuzzy decision tree based on the execution cost matrix, establish a distributed consensus coordination model with the conflict degree as the node weight, obtain the optimal execution timing by iteratively updating the node states, and generate control signals for each actuator according to the optimal execution timing, including:

[0038] Decompose the optimal control strategy based on the physical constraints of the controlled object and the dynamic characteristics of the actuator to obtain multiple control sub-strategies with execution dependencies, and construct an execution feature vector for each control sub-strategy that includes an execution time interval, resource type and its occupancy, task priority, and control objective importance;

[0039] Calculate the time dimension overlap degree, resource dimension competition degree, priority dimension difference degree, and control objective correlation degree between any two control sub-strategies based on the execution feature vector, and perform a weighted combination of the time dimension overlap degree, resource dimension competition degree, priority dimension difference degree, and control objective correlation degree using an adaptive weight to construct an execution cost matrix for the control sub-strategies, where each element of the execution cost matrix for the control sub-strategies represents the execution cost between the corresponding two control sub-strategies;

[0040] Input the execution cost matrix into a fuzzy decision tree with a multi-layer fuzzy rule base. The multi-layer fuzzy rule base includes a rule subset based on the time dimension overlap degree, a rule subset based on the resource dimension competition degree, a rule subset based on the priority dimension difference degree, and a rule subset based on the control objective correlation degree, and calculate the conflict degree between the control sub-strategies using a fuzzy inference mechanism;

[0041] Construct a distributed consensus coordination network with each control sub-strategy as a node. The node state includes the start time and duration of the execution time interval. Calculate the coupling strength between nodes based on the conflict degree, and use the normalized value of the coupling strength as the node weight, and the node weight is dynamically adjusted as the conflict degree changes;

[0042] Iteratively update the node states in the distributed consensus coordination network. The update amount of the node state is related to the state difference of adjacent nodes, the coupling strength, and the execution dependency, where the execution dependency is represented by a directed edge, until the distributed consensus coordination network reaches a consensus state to obtain the optimal execution timing that satisfies the timing constraints. According to the optimal execution timing, combined with the control dead zone, response delay, and saturation constraints of the actuator dynamic characteristics, generate control signals for each actuator.

[0043] In an alternative embodiment,

[0044] An adaptive dead - zone compensator is used to compensate the control signals of each actuator. Calculating the compensation gain based on the dead - zone characteristics of the actuator and correcting the control signal includes:

[0045] Collect the input characteristic curve and output characteristic curve of the actuator, extract the positive dead - zone threshold, negative dead - zone threshold and linear region slope in the input characteristic curve and output characteristic curve to obtain the actuator gain coefficient, establish an actuator dead - zone characteristic model, and set the compensation gain based on the average value of the positive dead - zone threshold and the negative dead - zone threshold;

[0046] Obtain the desired output and the current actuator output, and calculate the deviation between the desired output and the current actuator output to obtain the system tracking error;

[0047] Calculate the sign value of the input control signal, multiply the compensation gain by the sign value to obtain the compensation amount, and add the compensation amount to the input control signal to obtain the compensated control signal;

[0048] Multiply the absolute value of the system tracking error by the absolute value of the input control signal to obtain the amplitude term, calculate the product of the system tracking error and the input control signal and extract its sign to obtain the direction term, multiply the amplitude term by the direction term to obtain the adaptive update amount of the compensation gain, perform an integral operation on the adaptive update amount to obtain a new compensation gain value, and multiply the new compensation gain value by the sign value of the input control signal to obtain a new compensation amount;

[0049] Add the new compensation amount to the input control signal to correct the compensated control signal to obtain the final control signal.

[0050] In the second aspect of the embodiments of the present invention, a thermal power combustion optimization control system based on deep learning is provided, including:

[0051] A first unit for detecting the distribution drift degree of the operating data of the thermal power unit by using the maximum mean discrepancy criterion, identifying the abnormal scenario type based on the multi - kernel density estimation method, calculating the data sampling weight according to the distribution drift degree and the abnormal scenario type, sampling the operating data of the thermal power unit according to the data sampling weight to obtain real - time operation samples, calculating the dynamic quantization threshold of the real - time operation samples based on the mutual information maximization criterion, encoding the real - time operation samples by using the dynamic quantization threshold to obtain a feature encoding sequence, and inputting the feature encoding sequence into a recurrent neural network to extract time - series features to obtain a working condition feature vector;

[0052] A second unit, configured to calculate a system state evaluation value according to the operating condition feature vector, generate a set of candidate control actions based on the system state evaluation value, use an adversarial learning network to determine whether each control action in the set of candidate control actions meets the system operation constraint conditions, screen to obtain a set of valid control actions that meet the constraint conditions, calculate the policy gradient of each control action in the set of valid control actions, and optimize the set of valid control actions based on the policy gradient to obtain an optimal control strategy;

[0053] A third unit, configured to decompose the optimal control strategy into multiple control sub-strategies, construct an execution cost matrix for the control sub-strategies, calculate the conflict degree of the control sub-strategies using a fuzzy decision tree based on the execution cost matrix, establish a distributed consistency coordination model with the conflict degree as the node weight, obtain an optimal execution timing by iteratively updating the node state, generate control signals for each actuator according to the optimal execution timing, compensate the control signals for each actuator using an adaptive dead zone compensator, calculate a compensation gain based on the actuator dead zone characteristics and correct the control signals, and input the corrected control signals into the thermal power unit execution terminal to achieve combustion optimization control.

[0054] The third aspect of the embodiments of the present invention

[0055] Provided is an electronic device, including:

[0056] A processor;

[0057] A memory for storing instructions executable by the processor;

[0058] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0059] The fourth aspect of the embodiments of the present invention,

[0060] Provided is a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0061] In this embodiment, the maximum average difference criterion is adopted to detect data distribution drift, and the data sampling weight is calculated according to the drift degree and the type of abnormal scenario, effectively improving the adaptability and robustness of the model to complex working conditions. The combination of the dynamic quantization threshold and the recurrent neural network can extract the working condition features more accurately, thereby improving the control accuracy. The adversarial learning network is used to judge whether the control action meets the system operation constraints, screen out effective control actions, ensure the safety of the control strategy, and avoid the risks brought by violating the constraint conditions. By constructing a control sub-strategy execution cost matrix and a distributed consistency coordination model, the execution timing of the control sub-strategy is optimized, execution conflicts are reduced, thereby reducing the execution cost of the control strategy and improving the control efficiency. The adaptive dead zone compensator can effectively compensate for the dead zone characteristics of the actuator, further improving the stability and efficiency of the control system. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 is a schematic flowchart of the method for optimizing the combustion control of thermal power based on deep learning according to an embodiment of the present invention;

[0063] Figure 2 is a schematic structural diagram of the system for optimizing the combustion control of thermal power based on deep learning according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0065] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0066] Figure 1 is a schematic flowchart of the method for optimizing the combustion control of thermal power based on deep learning according to an embodiment of the present invention, as Figure 1 shown, the method includes:

[0067] S101. Detect the distribution drift degree of the operating data of the thermal power unit using the maximum mean discrepancy criterion, identify the abnormal scenario types based on the multi-kernel density estimation method, calculate the data sampling weights according to the distribution drift degree and abnormal scenario types, sample the operating data of the thermal power unit according to the data sampling weights to obtain real-time operating samples, calculate the dynamic quantization threshold of the real-time operating samples based on the mutual information maximization criterion, encode the real-time operating samples using the dynamic quantization threshold to obtain a feature encoding sequence, and input the feature encoding sequence into a recurrent neural network to extract temporal features to obtain a working condition feature vector;

[0068] S102. Calculate the system state evaluation value according to the working condition feature vector, generate a control action candidate set based on the system state evaluation value, use an adversarial learning network to judge whether each control action in the control action candidate set meets the system operation constraint conditions, screen to obtain an effective control action set that meets the constraint conditions, calculate the policy gradients of each control action in the effective control action set, and optimize the effective control action set based on the policy gradients to obtain an optimal control strategy;

[0069] S103. Decompose the optimal control strategy into multiple control sub-strategies, construct an execution cost matrix of the control sub-strategies, calculate the conflict degree of the control sub-strategies using a fuzzy decision tree based on the execution cost matrix, establish a distributed consistency coordination model with the conflict degree as the node weight, obtain the optimal execution time sequence by iteratively updating the node states, generate control signals for each actuator according to the optimal execution time sequence, compensate the control signals for each actuator using an adaptive dead zone compensator, calculate the compensation gain based on the actuator dead zone characteristics and correct the control signals, and input the corrected control signals into the execution terminal of the thermal power unit to achieve combustion optimization control.

[0070] Among them, the maximum mean discrepancy criterion is used to measure the difference degree between two data distributions. By comparing the distributions of historical data and current data, the distribution drift of the operating data is detected to identify potential working condition changes or abnormal situations. The multi-kernel density estimation method calculates the probability distribution of data by combining multiple kernel functions, identifies the abnormal scenario types under complex working conditions, and improves the recognition accuracy of abnormal working conditions.

[0071] Among them, the data sampling weight refers to the weight value assigned according to the distribution drift and anomaly type of the operation data, which is used to adjust the contributions of different data during the sampling process to obtain a more representative real-time operation sample. The control sub-strategy execution cost matrix describes the costs that may be generated during the execution of each control sub-strategy, including multi-dimensional information such as time, resource competition, and priority conflicts. The fuzzy decision tree calculates the conflict degree between control sub-strategies through fuzzy rules and is used to optimize the strategy execution order. The distributed consistency coordination model ensures that the execution order of each control sub-strategy reaches the global optimum through dynamic adjustment between nodes. The adaptive dead zone compensator is a device used to correct the control signal of the actuator, which can dynamically adjust the compensation amount to make up for the control deviation caused by actuator characteristics (such as dead zone, delay, etc.).

[0072] In an alternative implementation, the maximum mean discrepancy criterion is used to detect the distribution drift degree of the operation data of the thermal power unit, the anomaly scenario type is identified based on the multi-kernel density estimation method, the data sampling weight is calculated according to the distribution drift degree and the anomaly scenario type, and sampling the operation data of the thermal power unit according to the data sampling weight to obtain a real-time operation sample includes:

[0073] The operation data of the thermal power unit is divided into a historical sample set and a current sample set according to a time window. The feature mapping function in the reproducing kernel Hilbert space is used to map the historical sample set and the current sample set to obtain historical mapping features and current mapping features. Gaussian kernel functions are constructed to perform kernel transformation on the historical mapping features and the current mapping features respectively to obtain historical kernel transformation features and current kernel transformation features. The unbiased estimation method is used to calculate the maximum mean discrepancy distance between the historical kernel transformation features and the current kernel transformation features, and the exponential smoothing method is used to filter the maximum mean discrepancy distance to obtain the distribution drift degree;

[0074] The initial probability density function of the operating condition feature vector is constructed by using the kernel density estimation method. The benchmark bandwidth of the initial probability density function is determined based on the grid search and cross-validation methods. The probability distribution of the historical sample set is used as the prior distribution. The kernel density estimation method is used to construct the trial distribution of the current sample set. The local data density ratio is calculated according to the prior distribution and the trial distribution. The bandwidth of the initial probability density function is updated based on the local data density ratio to obtain the adaptive bandwidth. The adaptive bandwidth is substituted into the initial probability density function to obtain the updated probability density function. The updated probability density function is logarithmically transformed to obtain the anomaly metric value of the operating condition feature vector. The anomaly scenario type is determined based on the anomaly metric value;

[0075] Construct a weight calculation model based on radial basis functions. Input the distribution drift degree, anomaly metric value, and anomaly scenario type into the weight calculation model. Use the backpropagation algorithm to optimize the network parameters of the weight calculation model to obtain sampling weights. Based on the sampling weights, construct a probability distribution function for importance sampling, and use the Markov chain Monte Carlo method to extract thermal power unit operation data from the probability distribution function to obtain real-time operation samples.

[0076] Exemplarily, first, divide the thermal power unit operation data into a historical sample set and a current sample set according to a time window. For example, use one month's data as the historical sample set and the most recent day's data as the current sample set. Then, use a mathematical tool called the reproducing kernel Hilbert space to map the data of the historical sample set and the current sample set into a high-dimensional feature space. It can be understood as transforming the data from the original space to a feature space that is easier to distinguish. In the feature space, construct Gaussian kernel functions for the historical sample set and the current sample set respectively, and perform kernel transformation. It can be understood as smoothing the data to highlight the overall characteristics of the data distribution. After that, use an unbiased estimation method to calculate the maximum mean discrepancy distance between the historical kernel-transformed features and the current kernel-transformed features. This distance can measure the degree of difference between the distributions of the two data sets. To avoid the influence of instantaneous fluctuations on the results, use the exponential smoothing method to filter the maximum mean discrepancy distance, and finally obtain the distribution drift degree. For example, the parameter of the exponential smoothing method is set to 0.9, indicating that the drift degree at the current moment is mainly affected by the previous moment.

[0077] Next, use the kernel density estimation method to construct an initial probability density function for the operating condition feature vector. It can be understood as estimating an initial probability model based on the data distribution of the historical sample set. To determine the optimal parameters of this model, use the grid search and cross-validation method to determine the benchmark bandwidth of the initial probability density function. For example, set the search range of the bandwidth to 0.1 to 1, with an interval of 0.1, and use five-fold cross-validation to select the best bandwidth. Use the probability distribution of the historical sample set as the prior distribution, and use the kernel density estimation method to construct a trial distribution for the current sample set. Calculate the local data density ratio based on the prior distribution and the trial distribution. This ratio reflects the change in the local data density of the current sample set relative to the historical sample set. Update the bandwidth of the initial probability density function based on the local data density ratio to obtain an adaptive bandwidth. It can be understood as adjusting the parameters of the initial probability model according to the data distribution of the current sample set to make it more suitable for the current operating condition. Substitute the adaptive bandwidth into the initial probability density function to obtain an updated probability density function. Perform a logarithmic transformation on the updated probability density function to obtain the anomaly metric value of the operating condition feature vector. According to the anomaly metric value, for example, set a threshold to judge the anomaly scenario type. For example, if the anomaly metric value is greater than the threshold, it is judged as an anomaly scenario, otherwise it is a normal scenario.

[0078] Finally, a weight calculation model based on radial basis function is constructed. The inputs of this model are the degree of distribution drift, the anomaly metric value, and the anomaly scenario type. For example, these three metrics are used as the input features of the model. The backpropagation algorithm is used to optimize the network parameters of the weight calculation model to obtain the sampling weights. For example, the learning rate is set to 0.01 and the number of iterations is set to 1000 to optimize the model parameters. An importance sampling probability distribution function is constructed based on the sampling weights. It can be understood that the sampling probability of the samples is adjusted according to the importance of the samples. The Markov chain Monte Carlo method is used to extract the operating data of thermal power units from the probability distribution function to obtain real-time operating samples. For example, the length of the Markov chain is set to 10000, the burn-in period is set to 1000, and 100 samples are extracted as real-time operating samples.

[0079] In this embodiment, by considering the distribution drift and anomaly scenarios, the sampling weights are adaptively adjusted, making the extracted samples more representative and capable of better reflecting the real-time operating state of thermal power units. The model trained based on the real-time operating samples can better capture the changes in working conditions, thereby improving the prediction accuracy and generalization ability of the model. Based on the more accurate model prediction results, the control strategy can be optimized to improve the operating efficiency and stability of thermal power units.

[0080] In an alternative embodiment, calculating the dynamic quantization threshold of the real-time operating samples based on the mutual information maximization criterion, encoding the real-time operating samples using the dynamic quantization threshold to obtain a feature encoding sequence, and inputting the feature encoding sequence into a recursive neural network to extract temporal features to obtain a working condition feature vector includes:

[0081] Construct a mutual information function between the real-time operating samples and their quantization values. The mutual information function includes the entropy of the real-time operating samples and the conditional entropy under the given quantization values. The gradient ascent method is used to iteratively optimize the mutual information function. In each iteration, the quantization threshold is updated based on the learning rate and the mutual information gradient. When the change value of the mutual information function is less than the preset mutual information function convergence threshold, the optimal quantization threshold is obtained, and the optimal quantization threshold is used as the dynamic quantization threshold;

[0082] Set the encoding bits and quantization levels based on the dynamic quantization threshold. The number of encoding bits is adaptively determined according to the distribution characteristics of the operating data of thermal power units. The quantization levels are calculated by the minimum description length criterion. The encoding bits and the quantization levels are input into an encoding function to map the real-time operating samples to obtain a feature encoding sequence;

[0083] Divide the feature coding sequence into multiple subsequences according to the temporal complexity, input each subsequence into a long short-term memory network respectively, obtain the cell state through the collaborative calculation of the forget gate and the input gate, process the cell state through the output gate to obtain the hidden state, and merge the results after processing each subsequence to obtain the cell state sequence and the hidden state sequence;

[0084] Calculate the weights for the hidden state sequence and the cell state sequence respectively using the temporal attention mechanism and the feature attention mechanism, and perform weighted fusion on the state sequences based on the weights to obtain the operating condition feature vector;

[0085] Calculate the reconstruction error term, the sparse constraint term, and the temporal consistency constraint term of the operating condition feature vector, combine them to obtain the loss function, and optimize the loss function using the backpropagation algorithm with an adaptive learning rate. When the value of the loss function is less than the preset threshold, obtain the optimized operating condition feature vector.

[0086] Exemplarily, first, perform dynamic quantization on the real-time operation samples. Construct the mutual information function between the real-time operation samples and their quantization values. This function consists of two parts: the entropy of the real-time operation samples themselves and the conditional entropy under the given quantization values. Entropy reflects the uncertainty of the sample values, and conditional entropy reflects the remaining uncertainty of the sample values after knowing the quantization values. Mutual information indicates how much information about the sample values can be obtained through the quantization values. Use the gradient ascent method to iteratively optimize the mutual information function. In each iteration, update the quantization threshold according to the learning rate and the gradient of the mutual information function with respect to the quantization threshold. The learning rate controls the step size of the update, and the mutual information gradient indicates the direction of the update. When the change value of the mutual information function is less than the preset convergence threshold, it is considered that the optimal quantization threshold has been found and used as the dynamic quantization threshold. For example, assume that the real-time operation samples are a series of temperature values [100, 102, 98, 105, 99], the initial quantization threshold is 100, and the threshold is continuously adjusted through the gradient ascent method until an optimal threshold, such as 101, is found, so that the quantized values can retain the original temperature information to the greatest extent.

[0087] Then, encode the real-time operation samples based on the dynamic quantization threshold. Adaptively determine the number of encoding bits according to the distribution characteristics of the operating data of the thermal power unit. For example, if the value range of the data is large, more encoding bits are required to represent it. The quantization level is calculated through the minimum description length criterion. The goal of the minimum description length criterion is to describe the data with the shortest encoding length to avoid overfitting. Input the determined number of encoding bits and the quantization level into the encoding function to map the real-time operation samples into a feature coding sequence. For example, assume that the dynamic quantization threshold is 101, the number of encoding bits is 3, and the quantization level is 4. Then the temperature value 100 will be encoded as "001", and the temperature value 105 will be encoded as "100".

[0088] Next, the feature encoding sequence is input into a recurrent neural network to extract temporal features. First, the feature encoding sequence is divided into multiple subsequences according to the temporal complexity. For example, a long sequence can be divided into multiple short sequences according to the sampling frequency. Then, each subsequence is separately input into a long short-term memory network (LSTM). The LSTM controls the flow of information through forget gates, input gates, and output gates, and can effectively capture long-term dependencies in sequence data. The forget gate determines which information to discard, the input gate determines which new information to retain, and the output gate determines which information to output. The output of the LSTM includes a cell state and a hidden state, which respectively reflect the internal memory and the current output of the LSTM. The results after processing each subsequence are merged to obtain a cell state sequence and a hidden state sequence.

[0089] Finally, the state sequences are weighted and fused to obtain the operating condition feature vector. The time attention mechanism is used to calculate the weights for the hidden state sequence, and the feature attention mechanism is used to calculate the weights for the cell state sequence. The time attention mechanism focuses on the importance of different time steps, and the feature attention mechanism focuses on the importance of different features. Based on the calculated weights, the state sequences are weighted and fused to obtain the final operating condition feature vector. For example, if the hidden state at a certain time step contributes more to the operating condition feature, the weight of that time step will be larger.

[0090] To optimize the operating condition feature vector, a loss function is constructed. The loss function consists of three parts: a reconstruction error term, a sparsity constraint term, and a temporal consistency constraint term. The reconstruction error term measures the difference between the feature vector and the original data, the sparsity constraint term encourages the sparsification of the feature vector, and the temporal consistency constraint term encourages the feature vectors at adjacent time steps to be consistent. The backpropagation algorithm with an adaptive learning rate is used to optimize the loss function. The adaptive learning rate can dynamically adjust the learning rate according to the change of the loss function, accelerating the convergence speed. When the value of the loss function is less than a preset threshold, it is considered that the optimized operating condition feature vector is obtained.

[0091] In this embodiment, the dynamic quantization threshold is determined by the mutual information maximization criterion, which can better retain the information of the original data, thereby improving the accuracy of feature extraction. Using a recurrent neural network to extract temporal features can effectively capture long-term dependencies in sequence data and enhance the robustness of the features. Through dynamic quantization and encoding, the data dimension can be reduced, thereby reducing the computational complexity.

[0092] In an alternative embodiment, the system state evaluation value is calculated according to the operating condition feature vector, a set of candidate control actions is generated based on the system state evaluation value, and an adversarial learning network is used to determine whether each control action in the set of candidate control actions satisfies the system operation constraint conditions. The steps to screen and obtain the set of valid control actions that meet the constraint conditions include:

[0093] Construct a state evaluation model with a two-stream network structure. The two-stream network includes a feature extraction stream and a state mapping stream. Input the working condition feature vector into the feature extraction stream for non-linear transformation to obtain a feature representation, input the feature representation into the state mapping stream to obtain the system state, construct a double value function network, input the system state into the double value function network, and take the minimum value of the outputs of the two networks in the double value function network to obtain the system state evaluation value;

[0094] Based on the system state and the system state evaluation value, construct a hybrid strategy network, calculate the hybrid weight through the hybrid strategy network, generate the mean vector and covariance matrix of each component based on the hybrid weight and the system state evaluation value, and transform the standard normal distribution random sampling result with the mean vector and covariance matrix to obtain a set of candidate control actions;

[0095] Use the expert demonstration actions that meet the system constraints as real samples and the actions that do not meet the constraints obtained by random sampling as generated samples and input them into an adversarial discriminator with a residual network structure. The residual network structure includes an input layer and multiple residual blocks. Obtain the input layer features through the input layer. Each residual block performs two-layer transformation on the input features to obtain intermediate features and add them to the input features to obtain residual features. Stack the input layer features and the residual features and then obtain the real sample discrimination value and the generated sample discrimination value through the discriminator output layer;

[0096] Calculate the discriminator loss function value based on the real sample discrimination value, the generated sample discrimination value, and the gradient penalty term, and update the network parameters of the adversarial discriminator in an alternating optimization manner. Repeat the iteration until the preset number of iterations is satisfied to obtain a trained adversarial discriminator;

[0097] Combine the set of candidate control actions with the system state and input them into the trained adversarial discriminator to obtain the candidate action evaluation score. Map the candidate action evaluation score to obtain a discrimination threshold, and select the actions with an evaluation score greater than the discrimination threshold as the set of effective control actions that meet the system constraint conditions.

[0098] Exemplarily, first, construct a two-stream network model for evaluating the system state. This two-stream network includes a feature extraction stream and a state mapping stream. Input the working condition feature vector describing the current operating condition of the system into the feature extraction stream. The feature extraction stream will perform a series of non-linear transformations on the input working condition feature vector to extract more representative feature representations. Then, input the extracted feature representations into the state mapping stream, and finally output the state of the system. To more accurately evaluate the system state, construct a double value function network. Input the system state into this double value function network, and the network will output two values. Take the smaller value of these two values as the evaluation value of the system state.

[0099] Suppose the operating condition feature vector is a 3D vector [25, 1.2, 100] containing temperature, pressure, and flow rate. The feature extraction stream can be composed of two fully connected layers, each containing 64 neurons and using the ReLU activation function. The state mapping stream can be a fully connected layer containing 32 neurons, outputting a scalar value representing the system state. Suppose the two values output by the dual value function network are 0.8 and 0.7 respectively, then the system state evaluation value is 0.7.

[0100] Next, construct a hybrid policy network based on the system state and the system state evaluation value. This network calculates a set of hybrid weights according to the input system state and evaluation value. Using these hybrid weights and the system state evaluation value, the mean vector and covariance matrix of multiple control action components can be generated. Then, randomly sample a set of data from the standard normal distribution, and transform the sampling results with the calculated mean vector and covariance matrix to generate a set of control action candidate sets.

[0101] Suppose the system state is 0.7, the system state evaluation value is 0.7, and the hybrid weights calculated by the hybrid policy network are [0.3, 0.7]. Suppose two control action components are generated, the mean vector of the first component is [1, 2], and the covariance matrix is [[1, 0], [0, 1]]; the mean vector of the second component is [3, 4], and the covariance matrix is [[2, 0], [0, 2]]. Randomly sample a two-dimensional vector [0.5, -0.2] from the standard normal distribution. After transformation, a set of control action candidate sets can be obtained.

[0102] Then, it is necessary to determine which actions in the generated control action candidate sets satisfy the system operation constraint conditions. For this purpose, construct an adversarial discriminator with a residual network structure. Use the expert demonstration actions that satisfy the system constraints as real samples, and the actions that do not satisfy the constraints obtained by random sampling as generated samples, and input them into the adversarial discriminator for training. The residual network structure includes an input layer and multiple residual blocks. The input layer converts the input data into a feature representation. Each residual block performs two-layer transformation on the input features to obtain intermediate features, and adds the intermediate features to the input features to obtain residual features. After superimposing the input layer features and the residual features of all residual blocks, the real sample discriminant value and the generated sample discriminant value are obtained through the discriminator output layer.

[0103] Suppose the expert demonstration action is [1.5, 2.5], and the action that does not satisfy the constraint obtained by random sampling is [5, 6]. The residual network can include two residual blocks, each residual block contains two fully connected layers, each layer contains 64 neurons and uses the ReLU activation function. The discriminator output layer is a fully connected layer, outputting a scalar value as the discriminant result.

[0104] Calculate the discriminator loss function value through the real sample discrimination value, the generated sample discrimination value, and the gradient penalty term. Update the network parameters of the adversarial discriminator in an alternating optimization manner, and repeat the iteration until the preset number of iterations is satisfied to obtain the trained adversarial discriminator.

[0105] Finally, combine the control action candidate set with the system state and input it into the trained adversarial discriminator to obtain the candidate action evaluation score. Map the candidate action evaluation score to obtain the discrimination threshold. Filter out the control actions with an evaluation score greater than the discrimination threshold as the set of valid control actions that meet the system constraint conditions.

[0106] Assume that the candidate action evaluation scores are [0.8, 0.6, 0.9], and the discrimination threshold obtained after mapping is 0.7. Then the set of valid control actions that meet the constraint conditions is the control actions corresponding to scores greater than 0.7.

[0107] In this embodiment, by using the adversarial learning network to determine whether the control action meets the system constraint conditions, it is possible to effectively avoid generating actions that violate the constraints, thereby improving the safety of the control strategy and ensuring the stable operation of the system. By using the hybrid strategy network to generate the control action candidate set, the exploration ability of the control strategy can be improved, and the optimal control action can be found faster, thereby improving the control efficiency. By using the state evaluation model with a two-stream network structure, the system state can be evaluated more accurately, enabling the control strategy to better adapt to different working conditions and improving the robustness and generalization ability of the control strategy.

[0108] In an alternative embodiment, calculating the policy gradient of each control action in the set of valid control actions and optimizing the set of valid control actions based on the policy gradient to obtain the optimal control strategy includes:

[0109] Construct a feature extraction network, input the working condition feature vector into the feature extraction network to obtain a feature representation, input the feature representation into a pre-trained policy optimization model, calculate the execution probability distribution of each control action in the set of valid control actions, and calculate the state value estimate of each control action;

[0110] During the actual operation of the system, execute each control action and record the corresponding actual reward value, calculate the temporal difference error between the actual reward value and the state value estimate, and use the temporal difference error as the advantage function;

[0111] Calculate the policy gradient based on the advantage function, construct the Fisher information matrix using the policy gradient, analyze the curvature characteristics of the policy gradient in different directions through the Fisher information matrix, determine the parameter update step sizes in each direction according to the curvature characteristics, such that the step size in the direction with larger curvature is smaller and the step size in the direction with smaller curvature is larger, multiply the step sizes in each direction by the corresponding policy gradient components and accumulate to obtain the parameter update amount;

[0112] Update the parameters of the policy optimization model using the parameter update amount to obtain an optimized policy optimization model, input the effective control action set into the optimized policy optimization model for evaluation, and select the control action with the highest evaluation value from the effective control action set as the optimal control strategy according to the evaluation result.

[0113] Exemplarily, first, construct a feature extraction network. The role of this network is to transform the working condition feature vector describing the current state of the system into a feature representation more suitable for processing by the policy optimization model. For example, in a scenario of controlling a robotic arm, the working condition feature vector may include the angles and speeds of the joints of the robotic arm and the position information of the target object. The feature extraction network can be a multi-layer perceptron or a convolutional neural network, and through training, it learns to transform these original features into more abstract and representative feature representations.

[0114] Next, input the extracted feature representation into a pre-trained policy optimization model. This model can be a recurrent neural network or a reinforcement learning model, such as an Actor-Critic network. The model will output the execution probability distribution of each control action in the effective control action set and the state value estimate corresponding to each control action. For example, the effective control action set of the robotic arm may include "move up", "move down", "move left", "move right", etc. The policy optimization model will predict the probability of each action and the expected reward obtained after executing the action. Suppose the probability of "move up" is 0.6, the probability of "move down" is 0.2, the probabilities of other actions are 0.1 respectively, and the state value estimate of "move up" is 10, the state value estimate of "move down" is 5, and the state value estimates of other actions are 2 and 3 respectively.

[0115] During the actual operation of the system, select and execute a control action according to the probability distribution output by the policy optimization model, and record the actual reward value obtained after executing the action. For example, after selecting to execute "move up", the robotic arm successfully grasps the target object and obtains an actual reward value of 20.

[0116] Then, calculate the difference between the actual reward value and the previous state value estimate. This difference is called the temporal difference error. For example, if the actual reward value for "moving up" is 20 and the state value estimate is 10, then the temporal difference error is 10. Use this temporal difference error as the advantage function. The advantage function reflects the gap between the reward obtained from actually performing an action and the expected reward.

[0117] Next, calculate the policy gradient based on the advantage function. The policy gradient indicates how to adjust the parameters of the policy optimization model so that the model can select better control actions. At the same time, use the policy gradient to construct the Fisher information matrix. The Fisher information matrix describes the curvature characteristics of the policy gradient in different directions. By analyzing the Fisher information matrix, the parameter update step size in each direction can be determined. A direction with a larger curvature indicates that a small change in the parameter will have a greater impact on the policy, so a smaller step size is required; conversely, a direction with a smaller curvature requires a larger step size. Multiply the step size in each direction by the corresponding policy gradient component and accumulate them to obtain the final parameter update amount.

[0118] Update the parameters of the policy optimization model using the calculated parameter update amount to obtain an optimized policy optimization model. Input the set of valid control actions into the optimized policy optimization model for evaluation, and select the control action with the highest evaluation value as the optimal control strategy according to the evaluation results. For example, the updated policy optimization model predicts that the probability of "moving up" is 0.8, the probability of "moving down" is 0.1, the probabilities of other actions are 0.05 and 0.05 respectively, and the state value estimate of "moving up" is 15, the state value estimate of "moving down" is 4, and the state value estimates of other actions are 1 and 2 respectively. Since the evaluation value of "moving up" is the highest, "moving up" is selected as the optimal control strategy.

[0119] In this embodiment, by continuously optimizing the policy and selecting better control actions, the efficiency and stability of the control system can be improved, and the control task can be better completed. It can handle complex working condition characteristics and adjust the policy according to the actual operating conditions, thereby improving the adaptability and robustness of the control system. Through automated learning and optimizing the control policy, the dependence on manual experience and intervention can be reduced, and the control cost can be lowered.

[0120] In an alternative implementation, decompose the optimal control strategy into multiple control sub - strategies, construct an execution cost matrix for the control sub - strategies, calculate the conflict degree of the control sub - strategies using a fuzzy decision tree based on the execution cost matrix, establish a distributed consistency coordination model with the conflict degree as the node weight, obtain the optimal execution timing by iteratively updating the node state, and generate control signals for each actuator according to the optimal execution timing, including:

[0121] Decompose the optimal control strategy based on the physical constraints of the controlled object and the dynamic characteristics of the actuator to obtain multiple control sub-strategies with execution dependencies, and construct an execution feature vector for each control sub-strategy, including the execution time interval, resource type and its occupancy, task priority, and control target importance;

[0122] Calculate the time dimension overlap degree, resource dimension competition degree, priority dimension difference degree, and control target association degree between any two control sub-strategies based on the execution feature vector, and use adaptive weights to perform weighted combination on the time dimension overlap degree, resource dimension competition degree, priority dimension difference degree, and control target association degree to construct a control sub-strategy execution cost matrix, where each element of the control sub-strategy execution cost matrix represents the execution cost between the corresponding two control sub-strategies;

[0123] Input the execution cost matrix into a fuzzy decision tree with a multi-layer fuzzy rule base. The multi-layer fuzzy rule base includes a rule subset based on the time dimension overlap degree, a rule subset based on the resource dimension competition degree, a rule subset based on the priority dimension difference degree, and a rule subset based on the control target association degree, and use a fuzzy inference mechanism to calculate the conflict degree between control sub-strategies;

[0124] Construct a distributed consistency coordination network with each control sub-strategy as a node. The node state includes the start time and duration of the execution time interval. Calculate the coupling strength between nodes based on the conflict degree, and use the normalized value of the coupling strength as the node weight, and the node weight is dynamically adjusted as the conflict degree changes;

[0125] Iteratively update the node state in the distributed consistency coordination network. The update amount of the node state is related to the state difference of adjacent nodes, the coupling strength, and the execution dependency. The execution dependency is represented by a directed edge. Until the distributed consistency coordination network reaches a consistency state to obtain an optimal execution timing that satisfies the timing constraint, and generate control signals for each actuator according to the optimal execution timing, combined with the control dead zone, response delay, and saturation constraint of the actuator dynamic characteristics.

[0126] Exemplarily, it is first necessary to analyze the physical constraints of the controlled object and the dynamic characteristics of the actuator. Physical constraints include the hard limits during the operation of the controlled object, such as the maximum pressure-bearing capacity of the boiler, the minimum combustion temperature, etc. These constraints are used to guide the boundary conditions for the decomposition of the optimal control strategy. The dynamic characteristics of the actuator refer to the physical behaviors exhibited by the actuator during actual operation, such as response time, output range, action delay, etc. For example, the response characteristics of a valve may include an opening time of 2 seconds, a closing time of 3 seconds, and an output flow range of 0 - 100 units.

[0127] When decomposing the optimal control strategy, each control sub-strategy represents an independent operation unit. For example, if the optimal control strategy is to optimize the fuel input and air ratio, the possible control sub-strategies after decomposition may include adjusting the fuel flow rate and regulating the air flow rate. Each control sub-strategy is associated with an execution time interval that defines the start and end times of its execution. For example, the adjustment of the fuel flow rate may be completed within the 1st second to the 10th second, while the adjustment of the air flow rate may need to be carried out from the 5th second to the 15th second. The resource type and its occupancy represent the types of resources required and the consumption amount when the sub-strategy is executed. For example, the fuel adjustment requires 50 units / second of fuel, and the air adjustment requires a fan power of 10 units.

[0128] The task priority describes the importance of the sub-strategy. For example, in a boiler system, the control priority of the fuel input may be higher than that of the air conditioning because the fuel directly affects the heat output. The importance of the control objective represents the relevance of the sub-strategy to the overall optimization objective. For example, the importance of fuel control in directly determining the heat energy supply is high, while the contribution of fan regulation to system noise optimization may be low.

[0129] When calculating the correlation between control sub-strategies, the time dimension overlap degree represents the proportion of the overlapping execution time intervals of two sub-strategies. For example, if the time interval of one sub-strategy is from the 1st second to the 10th second and the other is from the 5th second to the 15th second, then their time dimension overlap degree is 50%. The resource dimension competition degree reflects the degree of demand conflict of two sub-strategies for the same resource. For example, if two sub-strategies need to use a fan at the same time, their competition degree will be high. The priority dimension difference degree represents the importance difference between sub-strategies. For example, the difference between the high-priority fuel control and the low-priority fan regulation is large. The control objective correlation degree describes the synergy of the sub-strategy to the overall objective. For example, the adjustment of fuel and air directly affects the combustion efficiency, and their correlation degree is high.

[0130] The execution cost matrix of the control sub-strategy is a two-dimensional array, and each element represents the execution cost between two sub-strategies. For example, if the time overlap degree between sub-strategy A and sub-strategy B is 50%, the resource competition degree is 30%, the priority difference degree is 20%, and the correlation degree is 40%, their execution cost is 40 after weighted by weights.

[0131] The fuzzy decision tree is a tool for processing the execution cost matrix. Its multi-layer rule base contains rule subsets of different dimensions. For example, the time overlap rule subset may include "high overlap requires delayed execution", and the resource competition rule subset may include "high competition requires priority allocation of important sub-strategies". The fuzzy inference mechanism synthesizes these rules to obtain the conflict degree between sub-strategies. For example, if the conflict degree between two sub-strategies is 0.7, it means that they need to be significantly adjusted to avoid interference.

[0132] The node state of the distributed consistency coordination network includes the start time and duration of the execution time interval. For example, the state of node A may be "start time: the 2nd second, duration: 8 seconds". The coupling strength is the relationship strength between nodes. For example, the coupling strength between two nodes is 0.8, indicating that their states are closely related. The normalized coupling strength is used as the node weight. For example, when the node weight is 0.9, it means that its adjustment amplitude requires a higher priority.

[0133] In the iterative update of the network, each adjustment of the node state refers to the state differences and dependencies of adjacent nodes. For example, if there is a dependency between node A and node B, the state update of node A needs to ensure that it starts after node B is completed. Finally, the distributed network reaches a consistent state and outputs an execution time sequence that satisfies all constraints. For example, node A starts at the 2nd second and node B starts at the 10th second.

[0134] When generating the control signal, the control dead zone of the actuator is considered. For example, a valve cannot act under a signal of 0 - 5 units, and at this time, it needs to be compensated to more than 6 units to act effectively. The response delay represents the actual effective time of the control signal. For example, a signal needs to be delayed by 0.5 seconds to execute after being input. The saturation constraint represents the operating range of the actuator. For example, the output signal shall not exceed 100 units.

[0135] Through the above process, a control signal that satisfies the physical constraints, dynamic characteristics, and timing requirements is finally generated to optimize the overall performance of the system.

[0136] In this embodiment, through the decomposition of the optimal control strategy and the construction of the cost matrix, the control process in the complex system can be refined, and the execution accuracy and efficiency can be improved; the fuzzy decision tree combines multi - dimensional conflict analysis to effectively solve the resource competition and priority conflicts between control sub - strategies, ensuring the coordination and robustness of the system under multi - task conditions; the distributed consistency coordination model dynamically adjusts the execution timing of the control strategy, ensuring the timeliness and global consistency of the execution, especially suitable for multi - control - node systems with complex dependencies; combined with the dynamic characteristics of the actuator and the control dead - zone compensation mechanism, the response performance of the control signal can be optimized, reducing delays and errors, and improving the stability and control accuracy of the system operation; the overall solution realizes the intelligent management of the complex control system through iterative optimization, ensuring the safe, economical, and efficient operation of the combustion process.

[0137] In an alternative implementation, an adaptive dead - zone compensator is used to compensate the control signals of each actuator. Calculating the compensation gain based on the actuator dead - zone characteristics and correcting the control signal includes:

[0138] Collect the input characteristic curve and output characteristic curve of the actuator, extract the positive dead zone threshold, negative dead zone threshold and linear region slope in the input characteristic curve and output characteristic curve to obtain the actuator gain coefficient, establish an actuator dead zone characteristic model, and set a compensation gain based on the average value of the positive dead zone threshold and the negative dead zone threshold;

[0139] Obtain the desired output and the current actuator output, and calculate the deviation between the desired output and the current actuator output to obtain the system tracking error;

[0140] Calculate the sign value of the input control signal, multiply the compensation gain by the sign value to obtain a compensation amount, and add the compensation amount to the input control signal to obtain a compensated control signal;

[0141] Multiply the absolute value of the system tracking error by the absolute value of the input control signal to obtain an amplitude term, calculate the product of the system tracking error and the input control signal and extract its sign to obtain a direction term, multiply the amplitude term by the direction term to obtain an adaptive update amount of the compensation gain, perform an integration operation on the adaptive update amount to obtain a new compensation gain value, and multiply the new compensation gain value by the sign value of the input control signal to obtain a new compensation amount;

[0142] Add the new compensation amount to the input control signal to correct the compensated control signal to obtain the final control signal.

[0143] Exemplarily, first, it is necessary to collect the input characteristic curve and output characteristic curve of the actuator. These two curves describe the output behavior of the actuator under different input signals. The input characteristic curve refers to the relationship between the control input (such as voltage, flow rate, etc.) and the actuator output (such as rotational speed, position, etc.), while the output characteristic curve reflects the response characteristic of the actuator.

[0144] In the analysis of the input and output characteristic curves, the key parameters are the dead zone threshold and the linear region slope. The positive dead zone threshold is the minimum input signal value when the input signal is positive and the actuator output cannot change; the negative dead zone threshold is the maximum input signal value when the input signal is negative and the actuator output cannot respond. The linear region slope represents the rate of change of the actuator output after the input signal exceeds the dead zone, usually reflected as a gain value. For example, in a certain hydraulic actuator, when the input voltage is between 0.2V and -0.3V, the output of the actuator cannot change. Once the voltage exceeds this range, the output will respond and change. In this way, the dead zone characteristic model of the actuator is determined.

[0145] Next, by analyzing the deviation between the desired output and the current actuator output, the system tracking error can be obtained. The desired output usually comes from the set target of the control system, while the actual output of the actuator is restricted by its physical characteristics, load, and control signals. For example, in a temperature control system, the desired output may be the set temperature, and the current actuator output is the actually measured temperature value. The difference between the two is the tracking error.

[0146] After obtaining the tracking error, the sign value of the input control signal needs to be calculated. The sign value refers to the directionality of the input signal, positive or negative. This is because the dead zone characteristic of the actuator usually has directionality, that is, the input signal shows different responses in different directions. After obtaining the sign value, multiply the compensation gain by the sign value to get the compensation amount. The compensation amount is to adjust the compensation gain to correct the deviation caused by the dead zone and correct the input control signal to make up for the part not responded to due to the dead zone.

[0147] The next step is to calculate the amplitude term and the direction term based on the system tracking error and the absolute value of the input signal. The amplitude term is the product of the absolute value of the tracking error and the absolute value of the input signal, reflecting the magnitude of the deviation. The direction term is the sign extracted from the product of the system tracking error and the input signal, indicating the direction of the deviation. Multiply the amplitude term and the direction term to obtain the adaptive update amount of the compensation gain. The goal of this step is to dynamically adjust the compensation gain according to the real-time error of the system so that it can adapt to different working states.

[0148] Finally, through the integral operation of the compensation gain update amount, a new compensation gain is obtained. The purpose of the integral operation is to gradually optimize the compensation gain over time to avoid excessive adjustments in the system instantaneously. At this time, based on the updated compensation gain value, it can be multiplied by the sign value of the input control signal to obtain the final compensation amount. Finally, add the compensation amount to the original control signal, and the corrected control signal is the final control output.

[0149] For example, in a hydraulic control system, the desired output temperature is 70°C, while the actual temperature is 72°C, resulting in an error of 2°C. The input voltage signal of the actuator is 5V. Suppose the compensation amount calculated through the aforementioned steps is 0.5V. Then, the corrected control signal will become 5.5V. Through this correction, the response of the actuator will be more accurate, and the system can reach the desired output more precisely.

[0150] In this embodiment, by dynamically calculating the compensation gain and adaptively adjusting the compensation strategy, precise control of the actuator under different operating states can be achieved, improving the robustness and adaptability of the system. Modeling using the input characteristic curve and output characteristic curve enables a comprehensive understanding of the actuator characteristics and targeted compensation, making the control signal more in line with the actual requirements. By adaptively updating the compensation amount, it is ensured that the compensation strategy is optimized with the change of the real-time system error, thereby enhancing the system's adaptability to complex working conditions and reducing the impact of non-linear factors on the control performance. Finally, this solution can provide efficient, precise, and stable control in various industrial scenarios, improving the performance of the actuator and extending its service life.

[0151] Figure 2 FIG. is a schematic structural diagram of an optimized control system for thermal power combustion based on deep learning according to an embodiment of the present invention, as Figure 2 shown, the system includes:

[0152] The first unit is used to detect the distribution drift degree of the operating data of the thermal power unit by using the maximum mean difference criterion, identify the abnormal scenario type based on the multi-core density estimation method, calculate the data sampling weight according to the distribution drift degree and abnormal scenario type, sample the operating data of the thermal power unit according to the data sampling weight to obtain real-time operating samples, calculate the dynamic quantization threshold of the real-time operating samples based on the mutual information maximization criterion, encode the real-time operating samples by using the dynamic quantization threshold to obtain a feature encoding sequence, and input the feature encoding sequence into a recurrent neural network to extract temporal features to obtain a working condition feature vector;

[0153] The second unit is used to calculate the system state evaluation value according to the working condition feature vector, generate a control action candidate set based on the system state evaluation value, use an adversarial learning network to judge whether each control action in the control action candidate set meets the system operation constraint conditions, screen to obtain an effective control action set that meets the constraint conditions, calculate the policy gradient of each control action in the effective control action set, and optimize the effective control action set based on the policy gradient to obtain an optimal control strategy;

[0154] The third unit is used to decompose the optimal control strategy into multiple control sub-strategies, construct an execution cost matrix of the control sub-strategies, calculate the conflict degree of the control sub-strategies by using a fuzzy decision tree based on the execution cost matrix, establish a distributed consistency coordination model with the conflict degree as the node weight, obtain the optimal execution time sequence by iteratively updating the node state, generate control signals for each actuator according to the optimal execution time sequence, compensate the control signals for each actuator by using an adaptive dead zone compensator, calculate the compensation gain based on the actuator dead zone characteristics and correct the control signals, and input the corrected control signals into the execution terminal of the thermal power unit to achieve combustion optimization control.

[0155] In a third aspect of the embodiments of the present invention,

[0156] a kind of electronic device is provided, including:

[0157] a processor;

[0158] a memory for storing instructions executable by the processor;

[0159] wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0160] In a fourth aspect of the embodiments of the present invention,

[0161] a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0162] The present invention may be a method, a device, a system and / or a computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are loaded.

[0163] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A thermal power combustion optimization control method based on deep learning, characterized in that, Including: Detect the distribution drift degree of the operating data of the thermal power unit by using the maximum mean discrepancy criterion, identify the types of abnormal scenarios based on the multi-kernel density estimation method, calculate the data sampling weights according to the distribution drift degree and the types of abnormal scenarios, sample the operating data of the thermal power unit according to the data sampling weights to obtain real-time operating samples, calculate the dynamic quantization threshold of the real-time operating samples based on the mutual information maximization criterion, encode the real-time operating samples by using the dynamic quantization threshold to obtain a feature encoding sequence, and input the feature encoding sequence into a recurrent neural network to extract temporal features to obtain a working condition feature vector; Calculate the system state evaluation value according to the working condition feature vector, generate a control action candidate set based on the system state evaluation value, use an adversarial learning network to judge whether each control action in the control action candidate set meets the system operation constraint conditions, screen to obtain an effective control action set that meets the constraint conditions, calculate the policy gradient of each control action in the effective control action set, and optimize the effective control action set based on the policy gradient to obtain an optimal control strategy; Decompose the optimal control strategy into multiple control sub-strategies, construct an execution cost matrix for the control sub-strategies, calculate the conflict degree of the control sub-strategies by using a fuzzy decision tree based on the execution cost matrix, establish a distributed consistency coordination model with the conflict degree as the node weight, obtain the optimal execution time sequence by iteratively updating the node state, generate control signals for each actuator according to the optimal execution time sequence, perform compensation processing on the control signals for each actuator by using an adaptive dead zone compensator, calculate the compensation gain based on the actuator dead zone characteristics and correct the control signals, and input the corrected control signals into the execution terminal of the thermal power unit to achieve combustion optimization control.

2. The method according to claim 1, wherein Detect the distribution drift degree of the operating data of the thermal power unit by using the maximum mean discrepancy criterion, identify the types of abnormal scenarios based on the multi-kernel density estimation method, calculate the data sampling weights according to the distribution drift degree and the types of abnormal scenarios, and sampling the operating data of the thermal power unit according to the data sampling weights to obtain real-time operating samples includes: Divide the operating data of the thermal power unit into a historical sample set and a current sample set according to a time window, map the historical sample set and the current sample set by using a feature mapping function in a reproducing kernel Hilbert space to obtain historical mapped features and current mapped features, construct Gaussian kernel functions to perform kernel transformation on the historical mapped features and the current mapped features respectively to obtain historical kernel transformation features and current kernel transformation features, calculate the maximum mean discrepancy distance between the historical kernel transformation features and the current kernel transformation features by using an unbiased estimation method, and filter the maximum mean discrepancy distance by using an exponential smoothing method to obtain the distribution drift degree; The initial probability density function of the operating condition feature vector is constructed by using the kernel density estimation method. The benchmark bandwidth of the initial probability density function is determined based on the grid search and cross-validation methods. The probability distribution of the historical sample set is used as the prior distribution. The trial distribution of the current sample set is constructed by using the kernel density estimation method. The local data density ratio is calculated according to the prior distribution and the trial distribution. The bandwidth of the initial probability density function is updated based on the local data density ratio to obtain the adaptive bandwidth. The adaptive bandwidth is substituted into the initial probability density function to obtain the updated probability density function. The updated probability density function is logarithmically transformed to obtain the anomaly metric value of the operating condition feature vector. The anomaly scenario type is determined based on the anomaly metric value; A weight calculation model based on the radial basis function is constructed. The distribution drift degree, the anomaly metric value, and the anomaly scenario type are jointly input into the weight calculation model. The network parameters of the weight calculation model are optimized by using the backpropagation algorithm to obtain the sampling weights. The probability distribution function of importance sampling is constructed based on the sampling weights. The Markov chain Monte Carlo method is used to extract the thermal power unit operation data from the probability distribution function to obtain the real-time operation samples.

3. The method according to claim 1, wherein Based on the mutual information maximization criterion, the dynamic quantization threshold of the real-time operation samples is calculated. The real-time operation samples are encoded by using the dynamic quantization threshold to obtain the feature coding sequence. The feature coding sequence is input into the recurrent neural network to extract the temporal features to obtain the operating condition feature vector, including: A mutual information function between the real-time operation samples and their quantization values is constructed. The mutual information function includes the entropy of the real-time operation samples and the conditional entropy under the given quantization values. The mutual information function is iteratively optimized by using the gradient ascent method. In each iteration, the quantization threshold is updated based on the learning rate and the mutual information gradient. When the change value of the mutual information function is less than the preset mutual information function convergence threshold, the optimal quantization threshold is obtained. The optimal quantization threshold is used as the dynamic quantization threshold; Based on the dynamic quantization threshold, the coding bits and the quantization levels are set. The number of coding bits is adaptively determined according to the distribution characteristics of the thermal power unit operation data. The quantization levels are calculated by using the minimum description length criterion. The coding bits and the quantization levels are input into the coding function to map the real-time operation samples to obtain the feature coding sequence; The feature coding sequence is divided into multiple subsequences according to the temporal complexity. Each subsequence is respectively input into the long short-term memory network. The cell state is obtained through the collaborative calculation of the forget gate and the input gate. The cell state is processed through the output gate to obtain the hidden state. The processed results of each subsequence are combined to obtain the cell state sequence and the hidden state sequence; Weights are calculated for the hidden state sequence and the cell state sequence respectively by using the temporal attention mechanism and the feature attention mechanism. The state sequences are weighted and fused based on the weights to obtain the operating condition feature vector; Calculate the reconstruction error term, sparse constraint term, and temporal consistency constraint term of the working condition feature vector, combine them to obtain a loss function, and use the backpropagation algorithm with an adaptive learning rate to optimize the loss function. When the value of the loss function is less than a preset threshold, the optimized working condition feature vector is obtained.

4. The method according to claim 1, wherein Calculate the system state evaluation value according to the working condition feature vector, generate a control action candidate set based on the system state evaluation value, and use an adversarial learning network to determine whether each control action in the control action candidate set satisfies the system operation constraint conditions. The filtered valid control action set that meets the constraint conditions includes: Construct a state evaluation model with a two-stream network structure. The two-stream network includes a feature extraction stream and a state mapping stream. Input the working condition feature vector into the feature extraction stream for non-linear transformation to obtain a feature representation, and input the feature representation into the state mapping stream to obtain the system state. Construct a double value function network, input the system state into the double value function network, and take the minimum value of the outputs of the two networks in the double value function network to obtain the system state evaluation value; Construct a hybrid strategy network based on the system state and the system state evaluation value. Calculate the hybrid weight through the hybrid strategy network, generate the mean vector and covariance matrix of each component based on the hybrid weight and the system state evaluation value, and transform the standard normal distribution random sampling result with the mean vector and covariance matrix to obtain the control action candidate set; Use the expert demonstration actions that meet the system constraints as real samples and the randomly sampled actions that do not meet the constraints as generated samples and input them into an adversarial discriminator with a residual network structure. The residual network structure includes an input layer and multiple residual blocks. Obtain the input layer features through the input layer. Each residual block performs two-layer transformation on the input features to obtain intermediate features and add them to the input features to obtain residual features. Stack the input layer features and the residual features and obtain the real sample discriminant value and the generated sample discriminant value through the discriminator output layer; Calculate the discriminator loss function value based on the real sample discriminant value, the generated sample discriminant value, and the gradient penalty term, and use the alternating optimization method to update the network parameters of the adversarial discriminator. Repeat the iteration until the preset number of iterations is satisfied to obtain the trained adversarial discriminator; Input the control action candidate set and the system state combination into the trained adversarial discriminator to obtain the candidate action evaluation score, map the candidate action evaluation score to obtain the discriminant threshold, and select the actions with an evaluation score greater than the discriminant threshold as the valid control action set that meets the system constraint conditions.

5. The method according to claim 1, wherein Calculate the policy gradient of each control action in the valid control action set, and optimize the valid control action set based on the policy gradient to obtain the optimal control strategy, including: Construct a feature extraction network, input the working condition feature vector into the feature extraction network to obtain a feature representation, input the feature representation into a pre-trained policy optimization model, calculate the execution probability distribution of each control action in the valid control action set, and calculate the state value estimation of each control action; During the actual operation of the system, each control action is executed and the corresponding actual reward value is recorded. The temporal difference error between the actual reward value and the state value estimate is calculated, and the temporal difference error is used as the advantage function. Based on the advantage function, the policy gradient is calculated, and the Fisher information matrix is constructed using the policy gradient. The curvature characteristics of the policy gradient in different directions are analyzed through the Fisher information matrix. According to the curvature characteristics, the parameter update step size in each direction is determined, such that the step size in the direction with a larger curvature is smaller and the step size in the direction with a smaller curvature is larger. The step size in each direction is multiplied by the corresponding policy gradient component and accumulated to obtain the parameter update amount. The parameters of the policy optimization model are updated using the parameter update amount to obtain an optimized policy optimization model. The effective control action set is input into the optimized policy optimization model for evaluation, and the control action with the highest evaluation value is selected from the effective control action set as the optimal control strategy according to the evaluation result.

6. The method according to claim 1, wherein The optimal control strategy is decomposed into multiple control sub-strategies, a control sub-strategy execution cost matrix is constructed, and the conflict degree of the control sub-strategies is calculated using a fuzzy decision tree based on the execution cost matrix. The conflict degree is used as the node weight to establish a distributed consistency coordination model. The optimal execution time sequence is obtained by iteratively updating the node state, and each actuator control signal is generated according to the optimal execution time sequence, including: Based on the physical constraints of the control object and the dynamic characteristics of the actuator, the optimal control strategy is decomposed to obtain multiple control sub-strategies with execution dependencies, and an execution feature vector including the execution time interval, resource type and its occupancy, task priority, and control target importance is constructed for each control sub-strategy. Based on the execution feature vector, the time dimension overlap degree, resource dimension competition degree, priority dimension difference degree, and control target association degree between any two control sub-strategies are calculated. An adaptive weight is used to perform a weighted combination of the time dimension overlap degree, resource dimension competition degree, priority dimension difference degree, and control target association degree to construct a control sub-strategy execution cost matrix. Each element of the control sub-strategy execution cost matrix represents the execution cost between the corresponding two control sub-strategies. The execution cost matrix is input into a fuzzy decision tree with a multi-layer fuzzy rule base. The multi-layer fuzzy rule base includes a rule subset based on the time dimension overlap degree, a rule subset based on the resource dimension competition degree, a rule subset based on the priority dimension difference degree, and a rule subset based on the control target association degree. A fuzzy inference mechanism is used to calculate the conflict degree between the control sub-strategies. A distributed consistency coordination network is constructed with each control sub-strategy as a node. The node state includes the start time and duration of the execution time interval. The coupling strength between the nodes is calculated based on the conflict degree, and the normalized value of the coupling strength is used as the node weight. The node weight is dynamically adjusted as the conflict degree changes. Iteratively update the node states in the distributed consistency coordination network. The update amount of the node states is related to the state difference of adjacent nodes, the coupling strength, and the execution dependency relationship, where the execution dependency relationship is represented by directed edges, until the distributed consistency coordination network reaches a consistent state to obtain an optimal execution timing that satisfies the timing constraints. According to the optimal execution timing, combined with the control dead zone, response delay, and saturation constraint of the actuator dynamic characteristics, generate the control signals for each actuator.

7. The method according to claim 1, wherein Use an adaptive dead zone compensator to compensate the control signals of each actuator. Calculating the compensation gain and correcting the control signals based on the actuator dead zone characteristics includes: Collect the input characteristic curve and output characteristic curve of the actuator, extract the positive dead zone threshold, negative dead zone threshold, and linear region slope in the input characteristic curve and output characteristic curve to obtain the actuator gain coefficient, establish an actuator dead zone characteristic model, and set the compensation gain based on the average value of the positive dead zone threshold and the negative dead zone threshold; Obtain the desired output and the current actuator output, and calculate the deviation between the desired output and the current actuator output to obtain the system tracking error; Calculate the sign value of the input control signal, multiply the compensation gain by the sign value to obtain the compensation amount, and add the compensation amount to the input control signal to obtain the compensated control signal; Multiply the absolute value of the system tracking error by the absolute value of the input control signal to obtain the amplitude term, calculate the product of the system tracking error and the input control signal and extract its sign to obtain the direction term, multiply the amplitude term by the direction term to obtain the adaptive update amount of the compensation gain, perform an integral operation on the adaptive update amount to obtain a new compensation gain value, and multiply the new compensation gain value by the sign value of the input control signal to obtain a new compensation amount; Add the new compensation amount to the input control signal to correct the compensated control signal to obtain the final control signal.

8. A thermal power combustion optimization control system based on deep learning, which is used to implement the method described in any one of the preceding claims 1-7, characterized in that, Including: The first unit is used to detect the distribution drift degree of the thermal power unit operation data by using the maximum mean discrepancy criterion, identify the abnormal scenario type based on the multi-kernel density estimation method, calculate the data sampling weight according to the distribution drift degree and the abnormal scenario type, sample the thermal power unit operation data according to the data sampling weight to obtain the real-time operation sample, calculate the dynamic quantization threshold of the real-time operation sample based on the mutual information maximization criterion, encode the real-time operation sample by using the dynamic quantization threshold to obtain the feature coding sequence, and input the feature coding sequence into the recurrent neural network to extract the time series features to obtain the working condition feature vector; The second unit is used to calculate the system state evaluation value according to the working condition feature vector, generate a control action candidate set based on the system state evaluation value, use the adversarial learning network to judge whether each control action in the control action candidate set meets the system operation constraint conditions, screen to obtain an effective control action set that meets the constraint conditions, calculate the policy gradient of each control action in the effective control action set, and optimize the effective control action set based on the policy gradient to obtain the optimal control strategy; The third unit is used to decompose the optimal control strategy into multiple control sub-strategies, construct an execution cost matrix of the control sub-strategies, calculate the conflict degree of the control sub-strategies by using a fuzzy decision tree based on the execution cost matrix, establish a distributed consistency coordination model with the conflict degree as the node weight, obtain the optimal execution timing by iteratively updating the node states, generate control signals for each actuator according to the optimal execution timing, compensate the control signals for each actuator by using an adaptive dead zone compensator, calculate the compensation gain based on the dead zone characteristics of the actuator and correct the control signals, and input the corrected control signals into the execution terminal of the thermal power unit to achieve combustion optimization control.

9. An electronic device, characterized in that, It includes: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images