Thermal power generating unit control parameter optimization method based on artificial intelligence

Abstract

The invention provides a thermal power generating unit control parameter optimization method based on artificial intelligence. The method comprises the steps of collecting time sequence data in the operation process of a thermal power generating unit; constructing a long short-term memory network and a physical information neural network of the thermal power generating unit, and training the long short-term memory network and the physical information neural network by adopting the time sequence data to obtain respective prediction performance; according to respective prediction performance, setting respective weight coefficients of the long short-term memory network and the physical information neural network to obtain a hybrid model of the long short-term memory network and the physical information neural network; the prediction performance of the hybrid model is compared with the prediction performance of the long short-term memory network and the prediction performance of the physical information neural network; and if the prediction performance of the hybrid model is better than the prediction performance of the long-short-term memory network and the prediction performance of the physical information neural network, the hybrid model is connected to a DCS system of the thermal power generating unit, and high-precision optimization of the control parameters of the thermal power generating unit is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control of thermal power units, and particularly to an optimization method for control parameters of thermal power units based on artificial intelligence. Background Art

[0002] As a core component of the power system, the operating efficiency and stability of thermal power units are crucial for the reliability and economy of power supply. However, the control methods of traditional thermal power units mainly rely on empirical models and manual operations, and there are problems such as insufficient control accuracy, poor adaptability, and limited optimization ability. These problems make it difficult for thermal power units to operate efficiently and stably in the face of complex operating condition changes.

[0003] In recent years, with the rapid development of artificial intelligence technology, its application in the field of thermal power unit control has gradually attracted attention. Although existing research has tried to introduce intelligent algorithms to optimize the unit control strategy, most of the existing technologies focus on the application of single intelligent algorithms and fail to fully combine the physical characteristics of thermal power units for systematic optimization. Therefore, there are still obvious deficiencies in the optimization of control parameters of thermal power units in the existing technologies. Summary of the Invention

[0004] The present invention provides an optimization method for control parameters of thermal power units based on artificial intelligence to solve the technical problem that the control parameters of thermal power units cannot be optimized with high precision in the existing technologies.

[0005] On the one hand, the present invention provides an optimization method for control parameters of thermal power units based on artificial intelligence, including: Collecting time series data during the operation of the thermal power unit; Constructing a long short-term memory network and a physics-informed neural network of the thermal power unit, and training the long short-term memory network and the physics-informed neural network respectively with the time series data to obtain their respective prediction performances; Setting the respective weight coefficients of the long short-term memory network and the physics-informed neural network according to the respective prediction performances to obtain a hybrid model of the long short-term memory network and the physics-informed neural network; Comparing the prediction performance of the hybrid model with the respective prediction performances of the long short-term memory network and the physics-informed neural network; If the prediction performance of the hybrid model is respectively better than the respective prediction performances of the long short-term memory network and the physics-informed neural network, connecting the hybrid model to the DCS system of the thermal power unit to generate the control parameter correction coefficient and the feedforward quantity correction coefficient of the PID controller in real time, and optimizing the PID controller.

[0006] An optimization method for control parameters of a thermal power unit based on artificial intelligence according to the present invention constructs a long short-term memory network for the thermal power unit, including: Construct a long short-term memory network for the thermal power unit; wherein, the long short-term memory network includes a plurality of long short-term memory units, and each long short-term memory unit includes an input gate, a forget gate and an output gate, which are used to capture the long-term dependence relationship between different control parameters in the time series data.

[0007] An optimization method for control parameters of a thermal power unit based on artificial intelligence according to the present invention constructs a physics-informed neural network for the thermal power unit, including: Construct a physics-informed neural network for the thermal power unit; wherein, the boiler energy balance equation, the steam turbine efficiency calculation method, and the maximum and minimum limits of each control parameter in different load sections are introduced into the physics-informed neural network.

[0008] An optimization method for control parameters of a thermal power unit based on artificial intelligence according to the present invention, the construction of the physics-informed neural network for the thermal power unit includes: The boiler energy balance equation is: Qr = Q1 + Q2; Where Qr is the heat input to the boiler, Q1 is the effectively utilized heat of the boiler, and Q2 is the heat loss due to flue gas; The steam turbine efficiency calculation method is: η = Wout / Qin; Where η is the thermal efficiency of the steam turbine, Wout is the mechanical power output by the steam turbine, and Qin is the heat absorbed by the steam turbine from the heat source; The maximum and minimum limits of each control parameter in different load sections are introduced as physical constraints into the physics-informed neural network.

[0009] An optimization method for control parameters of a thermal power unit based on artificial intelligence according to the present invention, setting the respective weight coefficients of the long short-term memory network and the physics-informed neural network according to their respective prediction performances to obtain a hybrid model of the long short-term memory network and the physics-informed neural network, including: Perform weighted fusion on the prediction results of the long short-term memory network and the physics-informed neural network to form a hybrid model; According to the prediction performances of the long short-term memory network and the physics-informed neural network, assign weight coefficients to their respective prediction results; wherein, the magnitude of the weight coefficient is proportional to the corresponding prediction performance; The hybrid model is represented by the following formula: ; Wherein, is the prediction result of the hybrid model, is the prediction result of the long short-term memory network, is the prediction result of the physics-informed neural network, and α is the weight coefficient of the long short-term memory network.

[0010] According to an artificial intelligence-based optimization method for control parameters of thermal power units provided by the present invention, comparing the prediction performance of the hybrid model with the prediction performance of the long short-term memory network and the physics-informed neural network respectively includes: Determining the evaluation indexes of the prediction results after weighted fusion and the evaluation indexes of the prediction results of the long short-term memory network and the physics-informed neural network respectively; wherein, the evaluation indexes include at least one of mean square error, root mean square error, mean absolute error, and coefficient of determination; Comparing the evaluation indexes of the prediction results after weighted fusion with the evaluation indexes of the prediction results of the long short-term memory network and the physics-informed neural network respectively.

[0011] According to an artificial intelligence-based optimization method for control parameters of thermal power units provided by the present invention, if the prediction performance of the hybrid model is respectively better than the prediction performance of the long short-term memory network and the physics-informed neural network, it further includes: Migrating the hybrid model to other target units, including: Performing standardization or normalization processing on the input features of the source unit and the target unit corresponding to the hybrid model to eliminate the dimension difference, and adjusting the output layer of the hybrid model according to the prediction requirements of the target unit; Minimizing the data distribution difference between the source unit and the target unit through the feature space mapping method; Verifying the migrated model using the verification data of the target unit.

[0012] According to an artificial intelligence-based optimization method for control parameters of thermal power units provided by the present invention, adjusting the output layer of the hybrid model according to the prediction requirements of the target unit includes: Mapping the input parameters of the hybrid model to the equivalent parameters of the target unit; If the prediction targets of the target unit and the source unit are different, replace the output layer of the hybrid model according to the prediction requirements of the target unit, and re-initialize the weights of the hybrid model; The minimizing the data distribution difference between the source unit and the target unit through the feature space mapping method includes: Adopting the maximum mean discrepancy algorithm to map the data of the source unit and the target unit to the shared feature space respectively, and minimizing the marginal distribution difference; If there are differences in the operating conditions of the target unit, use the pseudo-label generation technology to adjust the conditional distribution; Calculate the marginal distribution MMD distance and the conditional distribution MMD distance between the source unit and the target unit respectively, define the dynamic balance factor μ, adjust the optimization weights of the marginal distribution and the conditional distribution through an adaptive algorithm, and the final loss function is L = μDmarginal+(1−μ)Dconditional. Use the training set data of the source unit and the target unit to train the model, and update the model parameters through backpropagation until the final loss reaches the minimum; Use the verification data of the target unit to verify the migrated model, and adjust according to the verification results to optimize the model performance, including: Use the verification subset data of the target unit to verify the hybrid model with the MMD adaptation layer, adjust the weights of the hybrid model, and calculate the evaluation indicators; When all evaluation indicators are better than the set threshold, stop the operation; If the set threshold cannot be met within the set number of iterations, perform hierarchical adjustment.

[0013] According to an optimization method for control parameters of a thermal power unit based on artificial intelligence provided by the present invention, it further includes: In the process of constructing the hybrid model, introduce a multi-objective optimization algorithm to optimize multiple performance indicators of the thermal power unit simultaneously; The multi-objective optimization algorithm is based on the Pareto optimization theory. By constructing a multi-objective optimization model, compare the prediction results of the hybrid model with the expected values of multiple performance indicators, and dynamically adjust the weight coefficients to achieve the balanced optimization between multiple performance indicators; The optimization objective function of the multi-objective optimization model is: ; Among them, is the prediction result of the hybrid model, n is the number of performance indicators, w i is the weight coefficient of the i-th performance indicator, f i is the evaluation function of the i-th performance indicator, and the weight coefficient is dynamically adjusted according to the actual operation requirements of the thermal power unit.

[0014] According to an optimization method for control parameters of a thermal power unit based on artificial intelligence provided by the present invention, it further includes: Perform real-time adaptive update on the hybrid model to cope with the dynamic changes and uncertain factors during the operation of the thermal power unit; The real-time adaptive update method includes: During the operation of the thermal power unit, monitor the environmental parameters and the unit operation state parameters in real time; Dynamically adjust the input features and weight coefficients of the hybrid model according to the monitored parameter changes; Adopt an online learning algorithm to incrementally update the hybrid model using the real-time collected operation data to ensure that the prediction performance of the model is always in the optimal state.

[0015] The method for optimizing control parameters of a thermal power unit based on artificial intelligence provided by the present invention constructs a long short-term memory network and a physics-informed neural network for the thermal power unit, and uses time series data to train the long short-term memory network and the physics-informed neural network respectively to obtain their respective prediction performances. According to their respective prediction performances, set the respective weight coefficients of the long short-term memory network and the physics-informed neural network to obtain a hybrid model of the long short-term memory network and the physics-informed neural network. Connect the hybrid model to the DCS system of the thermal power unit to generate the control parameter correction coefficient and the feedforward quantity correction coefficient of the PID controller in real time, and optimize the PID controller, which can effectively improve the prediction accuracy and optimization ability of the thermal power unit control system, and enhance the operation efficiency, economy and stability of the unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0017] Figure 1 is a schematic flow chart of the method for optimizing control parameters of a thermal power unit based on artificial intelligence provided by an embodiment of the present invention; Figure 2 is a schematic structural diagram of the device for optimizing control parameters of a thermal power unit based on artificial intelligence provided by an embodiment of the present invention; Figure 3 is a schematic structural diagram of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention with reference to the drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0019] Figure 1 is a schematic flow chart of the method for optimizing control parameters of a thermal power unit based on artificial intelligence provided by an embodiment of the present invention.

[0020] See Figure 1 , the method for optimizing control parameters of thermal power units based on artificial intelligence may include the following steps.

[0021] Step 101: Collect time series data during the operation of the thermal power unit.

[0022] In this step, the time series data refers to a series of related parameter data recorded in chronological order during the operation of the thermal power unit. For example, the time series data may include load command, active power, main steam pressure, reheat steam pressure, main steam temperature, main steam flow rate, integrated valve position command, turbine control valve opening, reheat steam temperature, feed water flow rate, feed water temperature, fuel quantity, air volume, secondary air temperature, primary air pressure, primary air temperature, oxygen content, furnace negative pressure, deaerator water level, condenser water level, unit back pressure, superheat degree, furnace temperature, flue gas temperature, flue gas flow rate, nitrogen oxide concentration, mill outlet temperature, ambient temperature, etc. Those skilled in the art can understand the specific meanings of the above control parameters, and no more detailed explanations will be given here.

[0023] The time series data can also be normalized. Then, the normalized time series data is divided into multiple subsets such as a training data set, a parameter optimization data set, a weight adjustment data set, and a verification data set. The training data set, parameter optimization data set, weight adjustment data set, and verification data set can be used for model training, parameter optimization, weight adjustment, and model function verification respectively.

[0024] Step 102: Construct a long short-term memory network and a physics-informed neural network for the thermal power unit, and use the time series data to train the long short-term memory network and the physics-informed neural network respectively to obtain their respective prediction performances.

[0025] In this step, the training data set can be used to train the long short-term memory network (Long Short-Term Memory Network, abbreviated as LSTM) and the physics-informed neural network (Physics-Informed Neural Network, abbreviated as PINN) respectively. Then, the verification data set is used to verify the evaluation indexes (such as mean square error, root mean square error, mean absolute error, and coefficient of determination, etc.) of the long short-term memory network and the physics-informed neural network, and the respective prediction performances can be obtained according to the magnitudes of the evaluation indexes.

[0026] Step 103: Set the respective weight coefficients of the long short-term memory network and the physics-informed neural network according to their respective prediction performances to obtain a hybrid model of the long short-term memory network and the physics-informed neural network.

[0027] Specifically, step 103 may include: Fuse the prediction results of the long short-term memory network and the physics-informed neural network through weighting to form a hybrid model; According to the prediction performances of the long short-term memory network and the physics-informed neural network, assign weight coefficients to their respective prediction results; wherein, the magnitude of the weight coefficient is proportional to the corresponding prediction performance; The hybrid model is represented by the following formula (1): (1); wherein, is the prediction result of the hybrid model, is the prediction result of the long short-term memory network, is the prediction result of the physics-informed neural network, and α is the weight coefficient of the long short-term memory network.

[0028] Step 104: Compare the prediction performance of the hybrid model with the respective prediction performances of the long short-term memory network and the physics-informed neural network.

[0029] Specifically, Step 104 may include: Determine the evaluation metrics for the prediction result after weighted fusion and the respective prediction results of the long short-term memory network and the physics-informed neural network; wherein, the evaluation metrics include at least one of the mean square error (abbreviation: MSE, measuring the squared difference between the predicted value and the true value), the root mean square error (abbreviation: RMSE, the square root of MSE, with the same unit as the original data), the mean absolute error (abbreviation: MAE, the average of the absolute differences between the predicted value and the true value), and the coefficient of determination (abbreviation: R², measuring the fitting degree of the model to the data); Compare each evaluation metric of the prediction result after weighted fusion with the evaluation metrics of the respective prediction results of the long short-term memory network and the physics-informed neural network.

[0030] For example, compare the mean square error of the hybrid model with the respective mean square errors of the long short-term memory network and the physics-informed neural network, compare the root mean square error of the hybrid model with the respective root mean square errors of the long short-term memory network and the physics-informed neural network, compare the mean absolute error of the hybrid model with the respective mean absolute errors of the long short-term memory network and the physics-informed neural network, and compare the coefficient of determination of the hybrid model with the respective coefficients of determination of the long short-term memory network and the physics-informed neural network.

[0031] Step 105: If the prediction performance of the hybrid model is respectively superior to the respective prediction performances of the long short-term memory network and the physics-informed neural network, connect the hybrid model to the DCS system of the thermal power unit to generate the control parameter correction coefficient and the feedforward quantity correction coefficient of the PID controller in real time, and optimize the PID controller.

[0032] In this step, the prediction performance of the hybrid model is better than that of the long short-term memory network and the physics-informed neural network respectively, that is, each evaluation index of the hybrid model is better than the corresponding evaluation indexes of the long short-term memory network and the physics-informed neural network.

[0033] In this embodiment, a long short-term memory network and a physics-informed neural network of a thermal power unit are constructed, and the long short-term memory network and the physics-informed neural network are respectively trained with time series data to obtain their respective prediction performances. According to their respective prediction performances, the weight coefficients of the long short-term memory network and the physics-informed neural network are set to obtain a hybrid model of the long short-term memory network and the physics-informed neural network. The hybrid model is connected to the DCS system of the thermal power unit to generate the control parameter correction coefficient and the feedforward quantity correction coefficient of the PID controller in real time, and optimize the PID controller, which can effectively improve the prediction accuracy and optimization ability of the thermal power unit control system, and enhance the operation efficiency, economy and stability of the unit.

[0034] In an embodiment of this specification, constructing a long short-term memory network of a thermal power unit may include: Constructing a long short-term memory network of a thermal power unit; wherein, the long short-term memory network includes a plurality of long short-term memory units, and each long short-term memory unit includes an input gate, a forget gate and an output gate, which are used to capture the long-term dependencies between different control parameters in the time series data.

[0035] In this embodiment, for example, the working process of the long short-term memory unit is as follows: Forget gate: Assume that at a certain moment, the influence of the change in fuel quantity on the main steam pressure gradually weakens. The forget gate will decide to discard the fuel quantity information at earlier time points because the influence of this information on the current state is already small.

[0036] Input gate: The input gate will decide to write the fuel quantity information at the current moment into the cell state. For example, if the fuel quantity increases at the current moment, the input gate will record this change because it may affect the future main steam pressure.

[0037] Updating the cell state: Combining the results of the forget gate and the input gate, the cell state will be updated to include the latest fuel quantity information. The cell state is like a "memory unit" that stores information that affects the current and future states.

[0038] Output gate: The output gate will output relevant information according to the cell state for predicting the main steam pressure at the next time point. For example, if the cell state records an increase in fuel quantity, the output gate may predict that the main steam pressure will rise.

[0039] In this way, the long short-term memory unit can dynamically remember and update the key information in the time series, capturing the long-term dependencies between different control parameters. For example, the change in fuel quantity may affect the main steam pressure after a period of time, and the long short-term memory network can effectively capture this delay effect, thus providing a more accurate prediction for the optimal control of thermal power units.

[0040] In one embodiment of this specification, constructing a physics-informed neural network for a thermal power unit includes: Constructing a physics-informed neural network for a thermal power unit; wherein, the boiler energy balance equation, the calculation method of turbine efficiency, and the maximum and minimum limits of each control parameter in different load sections are introduced into the physics-informed neural network.

[0041] Specifically, constructing the physics-informed neural network for a thermal power unit may include: The boiler energy balance equation is shown as formula (2) below: Qr = Q1 + Q2 (2); Wherein, Qr is the heat input to the boiler, Q1 is the effectively utilized heat of the boiler, and Q2 is the heat loss due to flue gas; The calculation method of turbine efficiency is shown as formula (3) below: η = Wout / Qin (3); Wherein, η is the thermal efficiency of the turbine, Wout is the mechanical power output by the turbine, and Qin is the heat absorbed by the turbine from the heat source; The maximum and minimum limits of each control parameter in different load sections are introduced into the physics-informed neural network as physical constraints.

[0042] In addition, during the process of constructing the physics-informed neural network for a thermal power unit, the heat loss due to incomplete chemical combustion, the heat loss due to incomplete mechanical combustion, the heat loss of the boiler body, and the physical heat loss of ash can also be ignored. The main reason for ignoring these factors is to simplify the model, improve the calculation efficiency, and ensure the practicality and accuracy of the model. These secondary heat losses have a relatively small impact on the overall performance under normal operating conditions, so they can be ignored in the model to achieve more efficient and practical optimal control.

[0043] In one embodiment of this specification, if the prediction performance of the hybrid model is respectively better than the prediction performance of the long short-term memory network and the physics-informed neural network, it further includes: Migrating the hybrid model to other target units, including: Normalizing or standardizing the input features of the source unit and the target unit corresponding to the hybrid model to eliminate the dimensional difference, and adjusting the output layer of the hybrid model according to the prediction requirements of the target unit.

[0044] By means of the feature space mapping method, the data distribution difference between the source unit and the target unit is minimized.

[0045] The verified data of the target unit is used to verify the migrated model.

[0046] In this embodiment, migrating the hybrid model to other target units significantly improves the cross-unit applicability and generalization ability of the hybrid model, while reducing the development cost and improving the prediction accuracy and stability of the model on different units.

[0047] In an embodiment of this specification, adjusting the output layer of the hybrid model according to the prediction requirements of the target unit includes: Step 1: Map the input parameters of the hybrid model to the equivalent parameters of the target unit.

[0048] Specifically, during the model migration process, the input parameters of the source unit and the target unit may differ in name, unit, or physical meaning. Therefore, it is necessary to map the input parameters of the source unit to the equivalent parameters of the target unit. For example, if the input parameter of the source unit is "fuel command", and the input parameter of the target unit is "fuel consumption", then "fuel command" needs to be converted to "fuel consumption" to ensure that the model can correctly understand and process the input data on the target unit.

[0049] Step 2: If the prediction targets of the target unit and the source unit are different (there are differences in the output quantities), then replace the output layer of the hybrid model according to the prediction requirements of the target unit, and re-initialize the weights of the hybrid model.

[0050] Specifically, if the prediction target of the target unit is different from that of the source unit, for example, the source unit predicts "main steam pressure", while the target unit needs to predict "turbine efficiency", then the output layer of the hybrid model needs to be replaced to meet the prediction requirements of the target unit. Since the structure and function of the output layer have changed, it is necessary to re-initialize the weights to avoid a decline in model performance caused by weight mismatch, so that the model can adapt to the specific prediction requirements of the target unit and improve the applicability and accuracy of the model.

[0051] By means of the feature space mapping method, minimizing the data distribution difference between the source unit and the target unit includes: Step 1: Adopt the Maximum Mean Discrepancy (MMD) algorithm to map the data of the source unit and the target unit to the shared feature space respectively, and minimize the marginal distribution difference.

[0052] Specifically, even if the input parameters have been mapped to equivalent parameters, there may still be differences in the data distributions of the source unit and the target unit. Feature space mapping methods (such as the Maximum Mean Discrepancy (MMD) algorithm) can map the data of the source unit and the target unit to a shared feature space, thereby minimizing the distribution differences between them. The MMD algorithm is a method for measuring the differences between two data distributions. By calculating the mean differences of the data of the source unit and the target unit in the feature space, the MMD algorithm can minimize the marginal distribution differences between them, ensuring that the data of the source unit and the target unit have similar distributions in the feature space, thus improving the migration performance of the model.

[0053] Step 2: If there are differences in the operating conditions of the target unit, use pseudo-label generation technology to adjust the conditional distribution.

[0054] Specifically, if the operating conditions of the target unit are significantly different from those of the source unit. For example, the target unit uses a different combustion method (the source unit may use the traditional pulverized coal combustion method, while the target unit may use the circulating fluidized bed combustion method) or operates in a different load section (the source unit may mainly operate in the high load section, while the target unit may operate more in the medium and low load sections), then it is necessary to adjust the conditional distribution.

[0055] The pseudo-label generation technology can help the model better adapt to the operating conditions of the target unit by generating pseudo-label data for the target unit, as follows: Step 1: Select unlabeled data of the target unit: Select a portion of unlabeled data from the operating data of the target unit. This data contains the actual operating parameters of the target unit but does not have corresponding labels (i.e., the predicted target values).

[0056] Step 2: Generate pseudo-labels using the source unit model: Use the model trained on the source unit to predict the unlabeled data of the target unit and generate pseudo-labels. These pseudo-labels are the predicted values of the model for the data of the target unit based on the patterns of the source unit.

[0057] Step 3: Screen high-quality pseudo-labels: Since the prediction of the source unit model for the target unit may not be completely accurate, it is necessary to screen out high-quality pseudo-labels. The screening can be carried out through the following methods: Confidence screening: Select the data with a relatively high prediction confidence of the model as pseudo-labels. For example, select the data with a prediction probability greater than a certain threshold (such as 0.9).

[0058] Consistency screening: If there are multiple sets of unlabeled data, select the part where the prediction results of multiple sets of data are consistent as pseudo-labels.

[0059] Step 4: Add the pseudo-labeled data to the training set: Add the filtered pseudo-labeled data to the training set of the target unit to form a new training set.

[0060] Step 5: Retrain the model: Retrain the model using the new training set (including the actual label data and pseudo-labeled data of the target unit). In this way, the model can learn the operating rules of the target unit while retaining the useful information of the source unit model.

[0061] Example: Suppose we are migrating a control parameter optimization model of a thermal power unit trained on a source unit to a target unit. The source unit mainly operates in the high load section, while the target unit mainly operates in the medium and low load sections. The following are the specific application steps of the pseudo-label generation technology: Step 1: Select the unlabeled data of the target unit: Select a part of the unlabeled data from the operating data of the target unit. This data contains the operating parameters of the target unit in the medium and low load sections.

[0062] Step 2: Generate pseudo-labels using the source unit model: Use the model trained on the source unit to predict the unlabeled data of the target unit and generate pseudo-labels. For example, predict parameters such as the main steam pressure and turbine efficiency of the target unit in the medium and low load sections.

[0063] Step 3: Screen high-quality pseudo-labels: Through confidence screening, select the data with a relatively high prediction confidence of the model as pseudo-labels. For example, select the data with a prediction probability greater than 0.9.

[0064] Step 4: Add the pseudo-labeled data to the training set: Add the filtered pseudo-labeled data to the training set of the target unit to form a new training set.

[0065] Step 5: Retrain the model: Retrain the model using the new training set. In this way, the model can learn the operating rules of the target unit in the medium and low load sections while retaining the useful information of the source unit model.

[0066] Step 3: Calculate the marginal distribution MMD distance and conditional distribution MMD distance between the source unit and the target unit respectively, define the dynamic balance factor μ, adjust the optimization weights of the marginal distribution and conditional distribution through an adaptive algorithm, and the final loss function is L = μDmarginal+(1−μ)Dconditional. Use the training set data of the source unit and the target unit to train the model, and update the model parameters through backpropagation until the final loss reaches the minimum. Dmarginal represents the marginal distribution MMD distance, and Dconditional represents the conditional distribution MMD distance.

[0067] Specifically, the marginal distribution MMD distance (Dmarginal): measures the difference in the marginal distributions of the source unit and the target unit in the feature space. The marginal distribution refers to the distribution of data without considering the conditional variables. The conditional distribution MMD distance (Dconditional): measures the difference in the conditional distributions of the source unit and the target unit in the feature space. The conditional distribution refers to the distribution of data given the conditional variables. The dynamic balance factor μ is a parameter between 0 and 1, which is used to balance the weights of the marginal distribution MMD distance and the conditional distribution MMD distance in the final loss function. Its value can be dynamically adjusted according to the performance of the model on the source unit and the target unit.

[0068] Verify the migrated model using the validation data of the target unit, including: Step 1: Use the validation subset data of the target unit to verify the hybrid model with the MMD adaptation layer, adjust the weights of the hybrid model, and calculate the evaluation metrics.

[0069] Specifically, the role of the validation subset data: The validation subset data of the target unit is a part selected from the actual operation data of the target unit, which is used to evaluate the performance of the migrated hybrid model on the target unit. These data contain the operation parameters of the target unit under different working conditions and the corresponding output results, which can truly reflect the prediction accuracy and adaptability of the model on the target unit.

[0070] The purpose of weight adjustment: By adjusting the weights of the hybrid model, the performance of the model can be optimized to better adapt to the operating characteristics of the target unit. The weight adjustment is based on the difference between the prediction results of the validation subset data and the true values, aiming to reduce the prediction error and improve the prediction accuracy of the model.

[0071] Calculation of evaluation metrics: The evaluation metrics include mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²), etc. These metrics measure the prediction performance of the model from different perspectives: MSE: Measures the squared difference between the predicted value and the true value. The smaller the value, the smaller the prediction error.

[0072] RMSE: The square root of MSE, which is consistent with the unit of the original data and can intuitively reflect the average error size between the predicted value and the true value.

[0073] MAE: The average of the absolute differences between the predicted value and the true value, reflecting the average deviation degree between the predicted value and the true value.

[0074] R²: Measures the goodness of fit of the model to the data. The value ranges from 0 to 1, and the closer it is to 1, the better the model fitting effect.

[0075] Step 2: When all evaluation metrics are better than the set threshold, stop the operation.

[0076] The significance of setting the threshold: The threshold is pre-determined according to the performance requirements and actual operating conditions of the target unit, and is used to judge whether the model has achieved satisfactory prediction accuracy and performance level. When all evaluation metrics are better than the set threshold, it means that the performance of the model on the target unit has reached the expected goal, and further optimization and adjustment can be stopped.

[0077] The conditions for stopping the operation: Only when all evaluation metrics meet the conditions, that is, are better than the set threshold, will the operation be stopped. This is because if any one of the evaluation metrics does not meet the standard, it may mean that the model has deficiencies in some aspects, which may affect the performance and reliability of the model in practical applications. Only when all metrics meet the standard can it be ensured that the model has good overall performance on the target unit.

[0078] Step 3: If the set threshold cannot be met within the set number of iterations, perform hierarchical adjustment. For example, freeze the bottom layer of LSTM and adjust the top-layer hybrid model and MMD adaptation layer until the index requirements are met.

[0079] Specifically, the role of setting the number of iterations: Setting the number of iterations is to limit the time and computational resource consumption in the model optimization process. In practical applications, model optimization is an iterative process that may require multiple adjustments and verifications to achieve satisfactory performance. By setting the number of iterations, it is possible to avoid the model optimization process from falling into an infinite loop or over-optimization to a certain extent, and at the same time ensure that the model optimization process is completed within a reasonable time range.

[0080] Reasons for hierarchical adjustment: If not all evaluation metrics can be made better than the set threshold within the set number of iterations, it indicates that there may be problems with the current model structure or parameter adjustment strategy, and a more refined adjustment method needs to be adopted. Hierarchical adjustment is an effective strategy. By freezing the bottom layer of the LSTM and retaining its ability to extract general temporal features, while adjusting the top-layer hybrid model and MMD adaptation layer, the performance of the model can be optimized more targeted, enabling it to better adapt to the operating characteristics of the target unit.

[0081] Freezing the bottom layer of the LSTM: The bottom layer of the LSTM is responsible for extracting general temporal features in time series data, and these features have a certain degree of generality among different units. Freezing the bottom layer of the LSTM can avoid destroying the extraction ability of these general features during the adjustment process, while reducing the computational amount and optimization difficulty.

[0082] Adjusting the top-layer hybrid model and MMD adaptation layer: The top-layer hybrid model and MMD adaptation layer are parts of the model that are closely related to the operating characteristics of the target unit. By adjusting this part, it can better adapt to the specific working conditions and operating parameters of the target unit, thereby improving the prediction accuracy and performance of the model on the target unit.

[0083] Judgment criteria for meeting the index requirements: During the process of hierarchical adjustment, it is still necessary to calculate evaluation metrics to determine whether the performance of the model meets the requirements. Only when all evaluation metrics are better than the set threshold is the model adjustment considered successful and meets the index requirements. If the index requirements still cannot be met after multiple hierarchical adjustments, it may be necessary to reconsider the model structure or optimization strategy, or further analyze and process the data of the target unit to find a more suitable solution.

[0084] In an embodiment of this specification, the method for optimizing control parameters of a thermal power unit based on artificial intelligence further includes: During the process of constructing the hybrid model, a multi-objective optimization algorithm is introduced to simultaneously optimize multiple performance metrics of the thermal power unit, including but not limited to power generation efficiency, pollutant emissions, fuel consumption, and equipment life; The multi-objective optimization algorithm is based on the Pareto optimization theory. By constructing a multi-objective optimization model, the prediction results of the hybrid model are compared with the expected values of multiple performance metrics, and the weight coefficients are dynamically adjusted to achieve balanced optimization among multiple performance metrics; The optimization objective function of the multi-objective optimization model is shown as the following formula (4): (4); Among them, is the prediction result of the hybrid model, n is the number of performance metrics, w i is the weight coefficient of the i-th performance metric, fi is the evaluation function for the i-th performance index, and the weight coefficient is dynamically adjusted according to the actual operation requirements of the thermal power unit.

[0085] In this embodiment, traditional optimization methods usually only focus on the optimization of a single performance index, such as improving power generation efficiency or reducing fuel consumption. However, in actual operation, the performance optimization of thermal power units needs to comprehensively consider multiple aspects, such as power generation efficiency, pollutant emissions, fuel consumption, and equipment life. Introducing a multi-objective optimization algorithm can simultaneously optimize these interrelated and potentially conflicting performance indexes to achieve balanced optimization of the overall performance. The multi-objective optimization algorithm is based on the Pareto optimization theory. Pareto optimization is a classic multi-objective optimization method that can find the optimal balance point among multiple objectives, generate a set of Pareto optimal solutions, and these solutions will not deteriorate another objective while improving one objective.

[0086] The core of the multi-objective optimization model is to compare the prediction results of the hybrid model with the expected values of multiple performance indexes, and achieve balanced optimization among multiple performance indexes by dynamically adjusting the weight coefficient.

[0087] In an embodiment of this specification, the method for optimizing the control parameters of a thermal power unit based on artificial intelligence further includes: Performing real-time adaptive update on the hybrid model to cope with the dynamic changes and uncertain factors during the operation of the thermal power unit.

[0088] Specifically, the real-time adaptive update method includes: Step 1: During the operation of the thermal power unit, real-time monitor environmental parameters (such as temperature, humidity, atmospheric pressure) and unit operation state parameters (such as load change rate, equipment fault signal).

[0089] Step 2: According to the monitored parameter changes, dynamically adjust the input features and weight coefficients of the hybrid model.

[0090] For example, if a large load change rate is detected, the feature weight related to load change can be increased. If a certain parameter has a greater impact on the current operating condition, the weight corresponding to this parameter can be increased.

[0091] Step 3: Adopt an online learning algorithm to perform incremental update on the hybrid model using the real-time collected operation data to ensure that the prediction performance of the model is always in the optimal state.

[0092] Specifically, the online learning algorithm is based on the Bayesian update theory, and dynamically adjusts the model parameters by calculating the prior probability and posterior probability. Its update formula is shown as formula (5) below: (5); Where, Represents the updated model parameters; Represents the model parameters before the update; η is the learning rate, which is used to control the step size of parameter update; L is the loss function, and ∇L represents the gradient of the loss function with respect to the model parameters, indicating the change direction and change rate of the loss function under the current parameters.

[0093] In some other embodiments of this specification, if the prediction performance of the hybrid model is not better than that of the long short-term memory network and the physics-informed neural network respectively, the following steps are performed: Readjust the weight coefficients: According to the prediction performance of the long short-term memory network and the physics-informed neural network on the validation dataset, dynamically adjust the weight coefficient α so that the weight coefficient is proportional to the prediction performance of the model, and iteratively optimize the weight coefficient through an optimization algorithm to minimize the prediction error of the hybrid model; Improve the model structure: Adjust the network structures of the long short-term memory network and the physics-informed neural network, including adjusting the number of network layers, the number of units, activation functions, etc., and at the same time introduce regularization techniques to constrain the model complexity to improve the model's expressive ability and generalization ability; Retrain the model: Re-partition the training dataset, validation dataset, and test dataset to ensure that the data distribution is reasonable and representative, and then use the adjusted weight coefficients and network structures to retrain the long short-term memory network and the physics-informed neural network until the model converges to a better solution; Optimize the fusion method of the hybrid model: Try other fusion methods, such as stacking fusion, weighted average and maximum / minimum fusion, etc., to explore better fusion strategies; at the same time introduce a multi-objective optimization algorithm, based on the Pareto optimization theory, optimize multiple performance indicators simultaneously, and dynamically adjust the weight coefficients to achieve balanced optimization among multiple performance indicators; Model validation and fine-tuning: Use an independent validation dataset to validate the adjusted hybrid model to ensure the stability of the improvement of the model performance, and fine-tune the parameters of the hybrid model according to the validation results, such as adjusting the learning rate, optimizer algorithm, etc., to further optimize the model performance.

[0094] In some other embodiments of this specification, connect the hybrid model to the DCS system of the thermal power unit to generate the control parameter correction coefficient and the feedforward quantity correction coefficient of the PID controller in real time, and optimize the PID controller, including: I. Connect the hybrid model to the DCS system Interface development and data interaction: Develop a dedicated interface program to seamlessly connect the hybrid model with the DCS (Distributed Control System) of the thermal power unit. This interface program can real-time obtain parameter data during the operation of the thermal power unit from the DCS system, such as load commands, main steam pressure, turbine control valve opening, etc. These data are the basic inputs for the hybrid model to perform predictions and optimizations.

[0095] This interface program is also responsible for real-time transmitting the results such as the correction coefficients calculated by the hybrid model back to the DCS system for adjusting the PID controller. During the data transmission process, it is necessary to ensure the accuracy and real-time of the data to avoid control errors caused by data delay or errors.

[0096] II. Real-time generate the control parameter correction coefficient and feedforward quantity correction coefficient of the PID controller Data preprocessing and input: After the hybrid model obtains the real-time operation parameters of the thermal power unit from the DCS system, first preprocess these data. The preprocessing includes operations such as data cleaning (removing outliers, filling missing values, etc.) and normalization to ensure that the data input into the hybrid model meets the requirements during model training.

[0097] Input the preprocessed data into the hybrid model. The hybrid model combines the processing ability of the long short-term memory network (LSTM) for time series data and the constraint of the physics-informed neural network (PINN) on physical laws to predict the future operation state of the thermal power unit. For example, predict key parameters such as the main steam pressure and turbine output power at the next moment.

[0098] Correction coefficient calculation: According to the prediction results of the hybrid model, calculate the control parameter correction coefficient of the PID controller. The parameters of the PID controller mainly include the proportional coefficient (Kp), integral coefficient (Ki), and derivative coefficient (Kd). The correction coefficient is calculated based on the deviation between the future state predicted by the hybrid model and the current set value. For example, if the hybrid model predicts that the main steam pressure will be lower than the set value, then it is necessary to adjust the parameters of the PID controller to make the system respond faster and increase the main steam pressure.

[0099] At the same time, calculate the feedforward quantity correction coefficient. Feedforward control is a control method based on disturbances, which can adjust the system in advance according to known disturbances (such as changes in fuel quantity, changes in load commands, etc.). The hybrid model can correct the feedforward quantity according to the predicted disturbances to make the feedforward control more accurate. For example, when the load command suddenly increases, the hybrid model can adjust the feedforward quantity of the fuel quantity in advance according to the predicted load change trend to quickly adapt to the load change.

[0100] Dynamic adjustment of the correction coefficient: The correction coefficient is not fixed, but dynamically adjusted according to the real-time operating state of the thermal power unit and the prediction results of the hybrid model. For example, when the operating conditions of the thermal power unit change (such as switching from high load to low load), the hybrid model will re-predict and update the correction coefficient according to the new prediction results. This dynamic adjustment mechanism enables the PID controller and feedforward control to better adapt to the complex operating environment of the thermal power unit and improve the control accuracy.

[0101] III. Optimize the PID Controller Parameter adjustment and control strategy update: Apply the calculated control parameter correction coefficient and feedforward quantity correction coefficient to the PID controller. Adjust the proportional coefficient, integral coefficient, and differential coefficient of the PID controller according to the correction coefficient. For example, if the correction coefficient indicates that the proportional action needs to be increased to accelerate the response speed, then appropriately increase the proportional coefficient Kp.

[0102] At the same time, update the feedforward control strategy. Adjust parameters such as the gain of the feedforward control according to the feedforward quantity correction coefficient, so that the feedforward control can more accurately compensate for disturbances. Through this parameter adjustment and control strategy update, the PID controller can better adapt to the dynamic changes of the thermal power unit and improve the stability and response speed of the control system.

[0103] Performance monitoring and feedback optimization: After applying the corrected PID controller to the control of the thermal power unit, continuously monitor the performance of the control system. Evaluate the optimization effect of the PID controller by monitoring the actual operating parameters of the thermal power unit (such as the deviation between the actual value and the set value of the main steam pressure, the stability of the steam turbine output power, etc.).

[0104] If it is found that the control effect does not meet the expectations, for example, the fluctuation of the main steam pressure is still large, then the hybrid model can be feedback optimized according to the actual operating data. The feedback optimization includes operations such as adjusting the parameters of the hybrid model and retraining the hybrid model to improve the prediction accuracy of the hybrid model, so as to further optimize the parameter correction coefficient and feedforward quantity correction coefficient of the PID controller.

[0105] In some other embodiments of this specification, the method for optimizing the control parameters of a thermal power unit based on artificial intelligence further includes: Perform real-time adaptive update on the hybrid model to cope with the dynamic changes and uncertain factors during the operation of the thermal power unit; The real-time adaptive update method includes: During the operation of the thermal power unit, real-time monitor the environmental parameters and unit operation state parameters; According to the monitored parameter changes, dynamically adjust the input features and weight coefficients of the hybrid model; An online learning algorithm is adopted to incrementally update the hybrid model using the real-time collected operation data, so as to ensure that the prediction performance of the model is always in the optimal state; Furthermore, the real-time adaptive update method further includes the following steps: Adaptive feature selection: According to the real-time operation state of the thermal power unit, automatically select the features that have a greater impact on the current working condition as inputs. For example, when a large load change rate is detected, features related to load change (such as fuel quantity, turbine control valve opening, etc.) are preferentially selected.

[0106] Dynamic weight adjustment strategy: An adaptive weight adjustment strategy is adopted to dynamically adjust the weight coefficients according to the magnitude of the model prediction error. When the prediction error is large, increase the weights of the relevant features to improve the prediction accuracy of the model.

[0107] Incremental learning and model update: An incremental learning algorithm (such as online gradient descent or Bayesian update) is used to update the hybrid model in real time. At each update, only part of the model parameters are adjusted, reducing the computational amount and improving the update efficiency.

[0108] Anomaly detection and handling: Real-time monitor the operation data of the thermal power unit to detect abnormal data (such as sensor failures or data fluctuations). When abnormal data is detected, automatically switch to a backup model or adopt a default control strategy to ensure the stable operation of the system.

[0109] Model performance evaluation and feedback: Regularly evaluate the performance of the hybrid model, and calculate evaluation metrics (such as mean square error, root mean square error, mean absolute error, and coefficient of determination). According to the evaluation results, dynamically adjust the structure and parameters of the model to further optimize the model performance.

[0110] Based on the same general inventive concept, the present invention also protects an artificial intelligence-based thermal power unit control parameter optimization device, as Figure 2 shown, Figure 2 is a schematic structural diagram of the artificial intelligence-based thermal power unit control parameter optimization device provided by an embodiment of the present invention. The artificial intelligence-based thermal power unit control parameter optimization device provided by the present invention will be described below. The artificial intelligence-based thermal power unit control parameter optimization device described below can be mutually corresponding and referred to the artificial intelligence-based thermal power unit control parameter optimization method described above.

[0111] The artificial intelligence-based thermal power unit control parameter optimization device includes: A data collection module 201 for collecting time series data during the operation of the thermal power unit; The model construction module 202 is configured to construct a long short-term memory network and a physics-informed neural network for a thermal power unit, and use time series data to train the long short-term memory network and the physics-informed neural network respectively to obtain their respective prediction performances. The model mixing module 203 is configured to set the respective weight coefficients of the long short-term memory network and the physics-informed neural network according to their respective prediction performances to obtain a hybrid model of the long short-term memory network and the physics-informed neural network. The performance comparison module 204 is configured to compare the prediction performance of the hybrid model with the respective prediction performances of the long short-term memory network and the physics-informed neural network. The model access module 205 is configured to, if the prediction performance of the hybrid model is respectively better than the respective prediction performances of the long short-term memory network and the physics-informed neural network, access the hybrid model to the DCS system of the thermal power unit, generate the control parameter correction coefficient and the feedforward quantity correction coefficient of the PID controller in real time, and optimize the PID controller.

[0112] Figure 3 It is a schematic structural diagram of an electronic device provided by an embodiment of the present invention.

[0113] As Figure 3 shown, the electronic device may include: a processor 310, a communication interface 320, a memory 330, and a communication bus 340. Among them, the processor 310, the communication interface 320, and the memory 330 complete communication with each other through the communication bus 340. The processor 310 can call the logical instructions in the memory 330 to execute the method for optimizing the control parameters of the thermal power unit based on artificial intelligence.

[0114] In addition, when the logical instructions in the above-mentioned memory 330 are implemented in the form of a software functional unit and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention essentially or the part that contributes to the prior art or a part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk, or an optical disc that can store program codes.

[0115] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the optimization method for the control parameters of a thermal power unit based on artificial intelligence provided by the above-mentioned various methods.

[0116] In yet another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is implemented to execute the optimization method for the control parameters of a thermal power unit based on artificial intelligence provided by the above-mentioned various methods.

[0117] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative labor.

[0118] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc., and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0119] 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 them; 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 for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A control parameter optimization method for a thermal power unit based on artificial intelligence, characterized in that: include: Collect time series data during the operation of thermal power units; Constructing a long short-term memory network and a physical information neural network of the thermal power unit, and using the time series data to train the long short-term memory network and the physical information neural network respectively to obtain their respective prediction performances; According to the respective prediction performances, setting respective weight coefficients of the long short-term memory network and the physical information neural network to obtain a hybrid model of the long short-term memory network and the physical information neural network; Comparing the prediction performance of the hybrid model with the prediction performance of the long short-term memory network and the physical information neural network respectively; If the prediction performance of the hybrid model is better than the prediction performance of the long short-term memory network and the physical information neural network respectively, the hybrid model is connected to the DCS system of the thermal power unit to generate the control parameter correction coefficient and feedforward correction coefficient of the PID controller in real time, and optimize the PID controller.

2. The method for optimizing control parameters of a thermal power unit based on artificial intelligence according to claim 1, characterized in that: Construct a long short-term memory network for thermal power plants, including: Construct a long short-term memory network of a thermal power unit; wherein the long short-term memory network includes a plurality of long short-term memory units, each of which includes an input gate, a forget gate and an output gate, for capturing the long-term dependency between different control parameters in the time series data.

3. The method for optimizing control parameters of a thermal power unit based on artificial intelligence according to claim 1, characterized in that: Construct the physical information neural network of thermal power units, including: A physical information neural network of a thermal power unit is constructed; wherein the boiler energy balance equation, the turbine efficiency calculation method, and the maximum and minimum limits of various control parameters in different load sections are introduced into the physical information neural network.

4. The method for optimizing control parameters of a thermal power unit based on artificial intelligence according to claim 3 is characterized in that: The method of constructing a physical information neural network of a thermal power plant unit includes: The boiler energy balance equation is: Qr=Q1+Q2; Among them, Qr is the heat input to the boiler, Q1 is the effective heat utilization of the boiler, and Q2 is the heat loss of exhaust gas; The calculation method of steam turbine efficiency is: η = Wout / Qin; Where η is the thermal efficiency of the steam turbine, Wout is the mechanical power output by the steam turbine, and Qin is the heat absorbed by the steam turbine from the heat source; The maximum limit and the minimum limit of each control parameter in different load sections are introduced into the physical information neural network as physical constraints.

5. The method for optimizing control parameters of a thermal power unit based on artificial intelligence according to claim 1, characterized in that: The method of setting the weight coefficients of the long short-term memory network and the physical information neural network according to the respective prediction performances to obtain a hybrid model of the long short-term memory network and the physical information neural network includes: Performing weighted fusion on the prediction results of the long short-term memory network and the physical information neural network to form a hybrid model; According to the prediction performance of the long short-term memory network and the physical information neural network, weight coefficients are assigned to the respective prediction results; wherein the size of the weight coefficient is proportional to the corresponding prediction performance; The mixed model is expressed by the following formula: ； in, is the prediction result of the hybrid model, is the prediction result of the long short-term memory network, is the prediction result of the physical information neural network, and α is the weight coefficient of the long short-term memory network.

6. The artificial intelligence-based thermal power unit control parameter optimization method according to claim 1, characterized in that: The comparing the prediction performance of the hybrid model with the prediction performance of the long short-term memory network and the physical information neural network, respectively, comprises: Determine the evaluation index of the prediction result after weighted fusion and the evaluation index of the prediction results of the long short-term memory network and the physical information neural network respectively; wherein the evaluation index includes at least one of mean square error, root mean square error, mean absolute error and determination coefficient; The evaluation indicators of the prediction results after weighted fusion are compared with the evaluation indicators of the prediction results of the long short-term memory network and the physical information neural network respectively.

7. The method for optimizing control parameters of a thermal power unit based on artificial intelligence according to claim 1, characterized in that: If the prediction performance of the hybrid model is better than the prediction performance of the long short-term memory network and the physical information neural network, the method further includes: Migrate the hybrid model to other target units, including: Standardizing or normalizing the input features of the source unit and the target unit corresponding to the hybrid model to eliminate dimensional differences, and adjusting the output layer of the hybrid model according to the prediction requirements of the target unit; Minimize the data distribution difference between the source unit and the target unit through feature space mapping method; The migrated model is validated using the validation data of the target unit.

8. The artificial intelligence-based thermal power unit control parameter optimization method according to claim 7, characterized in that: The step of adjusting the output layer of the hybrid model according to the prediction requirements of the target unit includes: Mapping the input parameters of the hybrid model to equivalent parameters of the target unit; If the prediction targets of the target unit and the source unit are different, the output layer of the hybrid model is replaced according to the prediction requirements of the target unit, and the weights of the hybrid model are reinitialized; The method of minimizing the data distribution difference between the source unit and the target unit by the feature space mapping method includes: The maximum mean difference algorithm is used to map the data of the source unit and the target unit to the shared feature space respectively, minimizing the difference in marginal distribution; If there are differences in the operating conditions of the target units, the pseudo-label generation technology is used to adjust the conditional distribution; The marginal distribution MMD distance and conditional distribution MMD distance of the source unit and the target unit are calculated respectively, and the dynamic balance factor μ is defined. The optimization weights of the marginal distribution and conditional distribution are adjusted through the adaptive algorithm. The final loss function is L=μDmarginal+(1−μ)Dconditional. The model is trained using the training set data of the source unit and the target unit, and the model parameters are updated through back propagation until the final loss is minimized. The migrated model is validated using the validation data of the target unit and adjustments are made based on the validation results to optimize model performance, including: The hybrid model with the MMD adaptation layer is verified using the target unit verification subset data, the weights of the hybrid model are adjusted, and the evaluation indicators are calculated; When all evaluation indicators are better than the set threshold, stop the operation; If the set threshold cannot be met within the set number of iterations, hierarchical adjustment is performed.

9. The method for optimizing control parameters of a thermal power unit based on artificial intelligence according to claim 1, characterized in that: Also includes: In the process of constructing the hybrid model, a multi-objective optimization algorithm is introduced to simultaneously optimize multiple performance indicators of the thermal power unit; The multi-objective optimization algorithm is based on the Pareto optimization theory. By constructing a multi-objective optimization model, the prediction results of the hybrid model are compared with the expected values ​​of multiple performance indicators, and the weight coefficients are dynamically adjusted to achieve balanced optimization among multiple performance indicators. The optimization objective function of the multi-objective optimization model is: ； in, is the prediction result of the hybrid model, n is the number of performance indicators, and w i is the weight coefficient of the i-th performance indicator, f i is the evaluation function of the ith performance indicator, and the weight coefficient is dynamically adjusted according to the actual operation demand of the thermal power unit.

10. The artificial intelligence-based thermal power unit control parameter optimization method according to claim 1, characterized in that: Also includes: The hybrid model is adaptively updated in real time to cope with dynamic changes and uncertainties during the operation of the thermal power unit; The real-time adaptive updating method comprises: During the operation of thermal power units, real-time monitoring of environmental parameters and unit operating status parameters; Dynamically adjusting input features and weight coefficients of the hybrid model according to monitored parameter changes; An online learning algorithm is used to incrementally update the hybrid model using real-time collected operating data to ensure that the prediction performance of the model is always in an optimal state.

Images