A method and system for predicting heat load on primary side of thermal power station

Abstract

The present invention relates to the field of data processing technology, and discloses a method and system for predicting the heat load on the primary side of a thermal power station, which is used to improve the accuracy of the heat load prediction on the primary side of the thermal power station. The method includes: collecting historical operation data of a target thermal power station, preprocessing the historical operation data, and obtaining a processing data set; performing feature importance analysis and screening on the processing data set to obtain a target feature set; constructing an initial BP network according to the target feature set, optimizing the weight threshold of the initial BP network, and obtaining a target BP network; training the target BP network model through the processing data set to obtain a heat load prediction model; collecting the current operation data of the target thermal power station in real time, inputting the current operation data into the heat load prediction model to perform heat load prediction, and obtaining a heat load prediction curve; extracting heat load prediction data according to the heat load prediction curve, generating a heat load prediction report, and transmitting the heat load prediction report to a data display terminal.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and system for predicting the primary side heat load of a thermal power station. Background Art

[0002] As an efficient and environmentally friendly heating mode, the central heating system has been widely used in northern China. It can centrally supply heating and domestic hot water, which not only meets the basic living needs of residents, but also facilitates the centralized management and efficient use of energy. In order to ensure that the heating system continuously, economically and efficiently meets the actual needs of users, achieves dynamic balance between supply and demand and optimizes resource allocation, it is particularly important to accurately predict the heat load. At present, heat load prediction technologies such as statistical modeling methods and machine learning algorithms have gradually emerged and have been applied in practice. Among them, artificial neural networks have been widely used in the field of heat load prediction due to their advantages such as strong adaptability, good nonlinear mapping ability and strong parallel processing ability, and have achieved good results.

[0003] However, traditional artificial neural networks often have defects such as slow convergence, easy to fall into local optimal solutions, and sensitivity to initial weights during the training process, which affects the generalization ability and prediction accuracy of the model and restricts its effect in practical applications. In addition, heat load prediction itself is a complex modeling problem with multiple variables, nonlinearity, and strong coupling. It not only requires the reasonable selection of influencing factors as input variables, but also requires the optimization of the topological structure and training algorithm of the neural network to improve its mapping ability. How to improve the global optimization ability of the neural network itself and how to perform reasonable feature engineering are key technical issues that need to be solved urgently. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a method and system for predicting the thermal load on the primary side of a thermal power station, which are used to improve the accuracy of the thermal load prediction on the primary side of a thermal power station.

[0005] The present invention provides a method for predicting heat load on the primary side of a thermal power station, comprising: collecting historical operation data of a target thermal power station, and preprocessing the historical operation data to obtain a processed data set;

[0006] Performing feature importance analysis and screening on the processed data set to obtain a target feature set;

[0007] An initial BP network is constructed according to the target feature set, and a weight threshold of the initial BP network is optimized by a preset improved grey wolf algorithm to obtain a target BP network;

[0008] Performing model training on the target BP network by using the processed data set to obtain a heat load prediction model;

[0009] Collecting the current operation data of the target thermal power station in real time, and inputting the current operation data into the thermal load prediction model to perform thermal load prediction, and obtain a thermal load prediction curve;

[0010] The heat load prediction data is extracted according to the heat load prediction curve, a heat load prediction report is generated, and the heat load prediction report is transmitted to a preset data display terminal.

[0011] In the present invention, the step of collecting the historical operation data of the target thermal power station and preprocessing the historical operation data to obtain a processed data set includes:

[0012] Collecting data on the heating system of the target thermal power station to obtain historical operation data;

[0013] Performing time alignment and frequency unification processing on the historical operation data to obtain aligned historical operation data;

[0014] Interpolating missing values ​​on the aligned historical operation data to obtain non-missing historical operation data;

[0015] Performing outlier detection on the non-missing historical operation data to obtain historical operation data containing anomaly marks;

[0016] Performing outlier elimination processing on the historical operation data containing abnormal marks to obtain historical operation data without abnormalities;

[0017] Performing data smoothing processing on the non-abnormal historical operation data according to a preset data smoothing index to obtain smoothed historical operation data;

[0018] Performing data segmentation on the smoothed historical operation data to obtain training historical operation data and test historical operation data;

[0019] Based on a preset normalization interval, the training historical operation data and the test historical operation data are respectively subjected to data normalization processing to obtain a normalized training data set and a normalized test data set;

[0020] Merging the normalized training data set and the normalized test data set to obtain a normalized historical operation data set;

[0021] The normalized historical operation data set is subjected to data de-redundancy processing to obtain a processed data set.

[0022] In the present invention, the step of performing feature importance analysis and screening on the processed data set to obtain a target feature set includes:

[0023] Performing feature extraction on the processed data set to obtain an original feature set;

[0024] Performing correlation analysis on the original feature set, calculating the Pearson correlation coefficient between each original feature and the heat load, and obtaining a feature correlation matrix;

[0025] Extracting diagonal elements of the feature correlation matrix to obtain an autocorrelation coefficient vector of each original feature;

[0026] The autocorrelation coefficient vectors of the original features are screened according to a preset threshold value, and the features with a value less than the preset threshold value are removed to obtain a preliminary screening feature set;

[0027] Building a decision tree model based on the initial screening feature set, and analyzing the decision tree model through the processed data set to obtain the information gain ratio of each original feature;

[0028] Sorting the initial screening feature set according to the information gain ratio of each original feature to obtain a sorted feature importance list;

[0029] Select the first N features in the feature importance list as the candidate feature set, where N is a preset feature quantity threshold;

[0030] Recursively screening the candidate feature set by a recursive feature elimination algorithm until a preset number of iterations is met to obtain a preferred feature subset;

[0031] Similar features are merged on the preferred feature subset to obtain the target feature set.

[0032] In the present invention, the steps of constructing an initial BP network according to the target feature set and optimizing the weight threshold of the initial BP network by a preset improved grey wolf algorithm to obtain a target BP network include:

[0033] Matching the node parameter set of the initial BP network according to the target feature set, wherein the node parameter set includes: the number of input nodes, the number of output nodes and the number of hidden layer nodes;

[0034] Initializing the parameters of the initial BP network through the node parameter set to obtain an initialization parameter set, wherein the initialization parameter set includes: network connection weights and network threshold data;

[0035] The initialization parameter set is converted and encoded into gray wolf individuals by using the improved gray wolf algorithm to obtain an initial gray wolf population;

[0036] The initial gray wolf population is optimized by a preset chaotic mapping strategy to obtain an optimized gray wolf population;

[0037] Extracting the fitness value of each optimized gray wolf individual in the optimized gray wolf population to obtain the fitness value of each optimized gray wolf individual;

[0038] Adjusting the search step length of each optimized gray wolf individual in the optimized gray wolf population, and updating the gray wolf individual position of each optimized gray wolf individual according to the fitness value of each optimized gray wolf individual, to obtain a target gray wolf population;

[0039] Decoding each target gray wolf in the target gray wolf population to obtain a target network parameter set, wherein the target network parameter set includes: a target network connection weight and a target network threshold data;

[0040] The initial BP network is optimized by weight threshold according to the target network parameter set to obtain a target BP network.

[0041] In the present invention, the step of training the target BP network model by processing the data set to obtain a heat load prediction model includes:

[0042] Initializing the training parameters of the target BP network to obtain a training parameter set, wherein the training parameter set includes: a learning rate, a momentum factor, and a regularization coefficient;

[0043] Based on the training parameter set, inputting the normalized training data set into the target BP network for forward propagation to obtain predicted output data;

[0044] Performing loss gradient calculation on the normalized training data set and the predicted output data through a preset loss function to obtain gradient information;

[0045] According to the gradient information, the target BP network is updated by a back propagation algorithm to obtain an updated BP network;

[0046] Performing model generalization processing on the updated BP network to obtain an initial heat load prediction model;

[0047] The initial heat load prediction model is subjected to hyperparameter optimization processing to obtain the heat load prediction model.

[0048] In the present invention, the step of collecting the current operation data of the target thermal power station in real time and inputting the current operation data into the thermal load prediction model to perform thermal load prediction to obtain a thermal load prediction curve includes:

[0049] Collecting the current operation data of the target thermal power station in real time, wherein the current operation data includes: water supply temperature, return water temperature, primary network water supply flow and heat;

[0050] Performing data stream conversion on the current operation data to obtain a real-time operation data stream;

[0051] The real-time operation data stream is filtered by a Kalman filter algorithm to obtain a smooth real-time operation data stream;

[0052] Extracting data segments from the smooth real-time running data stream to obtain real-time running segment data;

[0053] Extracting operation features from the real-time operation segment data to obtain an operation feature vector set;

[0054] Inputting the operating characteristic vector set into the heat load prediction model for forward propagation to obtain a heat load prediction value set;

[0055] Curve fitting is performed on the heat load prediction value set to obtain the heat load prediction curve.

[0056] In the present invention, the step of extracting heat load prediction data according to the heat load prediction curve, generating a heat load prediction report, and transmitting the heat load prediction report to a preset data display terminal includes:

[0057] Perform discrete sampling of the predicted heat load curve to obtain discrete prediction data;

[0058] Performing piecewise linear regression fitting on the discretized prediction data to generate a local trend feature vector;

[0059] Performing statistical feature vector extraction on the local trend feature vector to obtain a global statistical feature vector;

[0060] Performing a fast Fourier transform on the global statistical feature vector to obtain a periodic feature vector;

[0061] Performing equal-length subsequence clustering on the periodic feature vector to obtain a pattern feature vector;

[0062] The heat load prediction report is generated according to the pattern feature vector, and the heat load prediction report is transmitted to a preset data display terminal.

[0063] The present invention also provides a thermal power station primary side heat load prediction system, comprising:

[0064] A collection module, used to collect historical operation data of a target thermal power station, and pre-process the historical operation data to obtain a processed data set;

[0065] A screening module is used to perform feature importance analysis and screening on the processed data set to obtain a target feature set;

[0066] A construction module is used to construct an initial BP network according to the target feature set, and optimize the weight threshold of the initial BP network through a preset improved grey wolf algorithm to obtain a target BP network;

[0067] A training module, used for performing model training on the target BP network through the processed data set to obtain a heat load prediction model;

[0068] A prediction module, used for collecting the current operation data of the target thermal power station in real time, and inputting the current operation data into the thermal load prediction model to perform thermal load prediction, so as to obtain a thermal load prediction curve;

[0069] The extraction module is used to extract heat load prediction data according to the heat load prediction curve, generate a heat load prediction report, and transmit the heat load prediction report to a preset data display terminal.

[0070] In the technical solution provided by the present invention, the link of data preprocessing is adopted, and the quality and consistency of the original data are effectively improved by means of time alignment, missing value interpolation, outlier processing, data smoothing and normalization, laying a good data foundation for subsequent modeling. Secondly, feature engineering technology is introduced, and the feature subset most relevant to the heat load is selected from a large number of influencing factors as the model input by using methods such as Pearson correlation coefficient analysis, decision tree and recursive feature elimination algorithm, avoiding the noise caused by redundant features, improving the representativeness and distinguishability of the features, and helping to improve the fitting and prediction capabilities of the model. In addition, the improved gray wolf optimization algorithm is organically combined with the BP neural network to play a complementary role in the advantages of the two. On the one hand, the gray wolf optimization algorithm has the characteristics of strong global optimization ability and fast convergence speed. By integrating various improvement measures such as chaos strategy population initialization, nonlinear adjustment convergence factor strategy and proportional weight dynamic adjustment strategy, it does not fall into the local optimal solution, and improves the robustness and adaptability of the algorithm. On the other hand, the BP neural network can effectively mine the complex internal laws and dependencies in the heat load data with its powerful nonlinear mapping ability. The combination of the two not only ensures that the algorithm has good global search capabilities, but also makes the model have stronger fitting capabilities, thereby improving prediction accuracy and stability.

[0071] At the same time, in the real-time prediction stage, the real-time operation data is processed and analyzed through technical means such as Kalman filtering, curve fitting, and feature extraction, which can not only capture the changing trend of heat load in a timely manner, but also identify mutation points and abnormal situations, so as to achieve continuous and accurate prediction and provide strong data support for heating scheduling decisions. Finally, an intuitive heat load forecast report is generated to facilitate heating management personnel to understand the changes in heat load, formulate corresponding scheduling strategies, and achieve accurate matching of supply and demand and optimal resource allocation. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0073] Figure 1 This is a flow chart of a method for predicting heat load on the primary side of a thermal power station in an embodiment of the present invention;

[0074] Figure 2 Schematic diagram of a thermal load prediction system for the primary side of a thermal power station in an embodiment of the present invention. DETAILED DESCRIPTION

[0075] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0076] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0077] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0078] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , Figure 1 is a flow chart of a method for predicting heat load on the primary side of a thermal power station according to an embodiment of the present invention. Figure 1 As shown, the following steps are included:

[0079] S101, collecting historical operation data of the target thermal power station, and preprocessing the historical operation data to obtain a processed data set;

[0080] S102, performing feature importance analysis and screening on the processed data set to obtain a target feature set;

[0081] S103, constructing an initial BP network according to the target feature set, and optimizing the weight threshold of the initial BP network by using a preset improved grey wolf algorithm to obtain a target BP network;

[0082] S104, training the target BP network model by processing the data set to obtain a heat load prediction model;

[0083] S105, collecting the current operation data of the target thermal power station in real time, and inputting the current operation data into a thermal load prediction model to perform thermal load prediction, and obtain a thermal load prediction curve;

[0084] S106. Extract heat load prediction data according to the heat load prediction curve, generate a heat load prediction report, and transmit the heat load prediction report to a preset data display terminal.

[0085] It should be noted that the historical operation data of the target thermal power station is collected, which usually includes parameters such as water supply temperature, return water temperature, primary network water supply flow, heat, etc. However, due to communication failures, collection system failures, etc., the original data may be missing or abnormal. Therefore, it is necessary to pre-process the historical operation data to ensure data quality.

[0086] Specifically, the preprocessing link first standardizes the operating data from different sources in terms of time and frequency through time alignment and frequency unification to eliminate data inconsistency caused by sampling frequency differences. Taking a heating station in a certain community as an example, since the sampling frequency of the primary network water supply flow is 1 minute, and the sampling frequency of the supply and return water temperature is 5 minutes, the two need to be processed in a unified frequency. Next, for the problem of missing data, the scheme uses the method of calculating the average value of adjacent data for interpolation to retain the integrity and continuity of the data as much as possible. Taking the above-mentioned community heating station as an example, if the primary network water supply flow data is missing due to communication interruption at a certain time point, the missing value can be estimated and filled by the average value of the adjacent data before and after the time point. In addition, due to equipment failure or extreme weather conditions, the data may have outliers. The scheme uses an outlier detection method based on the 3σ principle to identify and process them. Taking a certain community as an example, if the water supply temperature data at a certain time point deviates from the normal level by more than 3 standard deviations due to extreme weather in the cold winter, the data will be marked as an outlier and eliminated.

[0087] For the data after removing outliers, the solution will also perform data smoothing to eliminate the impact of high-frequency noise and improve data availability. For example, moving average smoothing of the primary water supply flow data of a community can effectively suppress high-frequency noise caused by pipe network pressure fluctuations. After the above series of preprocessing steps, a processed high-quality data set can be obtained, laying the foundation for subsequent model construction and training.

[0088] After obtaining the processed data set, the Pearson correlation coefficient method, decision tree and recursive feature elimination algorithm are used to analyze and screen the data set for feature importance, and obtain the target feature set most relevant to the heat load. Taking a certain community as an example, it is found through analysis that the four features of primary side heating, primary side temperature recovery, heat load in the first two hours and heat load in the first hour have a strong linear correlation with the heat load at the current moment, so they can be used as input features to build a prediction model. Reasonable feature selection can not only reduce the noise interference caused by redundant features, but also improve the generalization ability of the model. After determining the target feature set, the scheme matches the number of input layer nodes of the BP neural network according to the number of features, initializes the network connection weights and threshold parameters, and constructs the initial BP network. Subsequently, the improved gray wolf optimization algorithm is applied to the training process of the initial BP network. Taking a certain community as an example, if the number of input features is 4, a BP network with 4 input nodes and 1 output node is initialized, and the network connection weights and threshold parameters are globally optimized by integrating the gray wolf algorithm with optimization strategies such as chaotic mapping, nonlinear convergence factor and dynamic weight adjustment to obtain a target BP network structure with better performance.

[0089] In the model training phase, the processed data set is input into the optimized target BP network, and the network parameters are continuously optimized through forward propagation, back propagation and parameter update steps, and finally a heat load prediction model with high accuracy is obtained. Taking a certain community as an example, the pre-processed historical operation data can be input into the optimized BP network as a training set, and a prediction model that reasonably fits the heat load pattern of the community is obtained after multiple iterations.

[0090] In actual applications, the current operating data of the target thermal power station, such as water supply temperature and return water temperature, will be collected in real time, and these data will be input into the trained prediction model for forward propagation calculation to obtain the predicted value of the heat load in the future. Taking a certain community as an example, the current supply and return water temperature of the community, the heat load in the previous two hours and other data can be obtained in real time. After inputting into the prediction model, the hourly heat load prediction curve in the next 24 hours can be obtained.

[0091] Finally, the forecast curve will be analyzed and extracted to generate a heat load forecast report, which will be transmitted to the preset data display terminal to present the heat supply management personnel in an intuitive visual form, providing a decision-making basis for them to formulate refined scheduling strategies. For example, by analyzing the heat load forecast curve using methods such as curve fitting and periodic analysis, the changing trend and periodic characteristics of the heat load in different time periods can be found, and this information can be presented in the form of charts or reports to provide scientific guidance for heat supply scheduling.

[0092] Specifically, regarding data preprocessing, in the actual heating process, the primary side heat load prediction results are affected by many complex factors, which involve not only the characteristics and operating status of the heating system itself, but also are closely related to external environmental conditions. The interaction of these factors makes the prediction of heating volume complex and changeable. Therefore, it is necessary to comprehensively consider multiple factors in the prediction process to ensure the accuracy and reliability of the prediction results. In order to solve the problem of outliers in the data, this application adopts an outlier detection method to ensure the accuracy of the data. The heating load is a seasonal load, and the load volume follows a normal distribution [8]. Therefore, the use of In principle, if the deviation between a value and the mean value exceeds 3 times the standard deviation, then this value is considered an outlier and is treated as a missing value. In order to eliminate the potential impact of different dimensions on algorithm performance, the data is normalized. In statistics, data normalization is to limit the data to a certain range through a certain algorithm, which not only facilitates subsequent data processing, but also improves the convergence speed of program operation. This application maps the data to between 0 and 1, and the normalized expression is shown in Formula 1: (1)

[0093] In the formula, y is the normalized data; x is the data before normalization; is the maximum value of the original data; is the minimum value of the original data.

[0094] Input characteristic variables, the Pearson correlation coefficient is mainly used to quantitatively evaluate the degree of linear correlation between two continuous variables. By calculating the Pearson correlation coefficient, we can intuitively understand the linear relationship between each influencing factor and the heat load, so as to determine the input characteristic variables of the model. The value range of the Pearson correlation coefficient is between -1 and 1, where 1 indicates a complete positive correlation, -1 indicates a complete negative correlation, and 0 indicates no linear correlation. Its expression is shown in Formula 2.

[0095] (2)

[0096] Where: is the correlation coefficient between variable x and variable y; is the i-th sample of variable x; is the variable y i samples; is the mean value of variable x; is the mean value of variable y; is the number of samples;

[0097] The Pearson correlation coefficient method is used to calculate and analyze the actual data, and the correlation coefficients of each eigenvector and the heat load are shown in Table 1. As can be seen from Table 1, the correlation coefficients of primary side heating, primary side temperature return, heat load in the first two hours, and heat load in the first hour are greater than 0.6, which are positively correlated with the heat load, primary side back pressure, primary side valve position feedback, and primary side instantaneous flow are weakly correlated with the load, and primary side supply pressure is negatively correlated with the load. Therefore, primary side heating, primary side temperature return, heat load in the first two hours, and heat load in the first hour are selected as input characteristic variables.

[0098] Table 1 Pearson correlation coefficient

[0099]

[0100] The Gray Wolf Algorithm simulates the predation behavior of gray wolf groups in nature and cleverly uses the mechanism of wolf group collaboration to achieve an optimal solution to the problem. The algorithm divides wolf groups into four levels: Layer wolf pack, Layer wolf pack, Layer wolf pack and Layer wolf pack, The stratum wolf is the leader of the wolf pack and is responsible for leading the entire wolf pack to hunt; Layer Wolf Pack Assist Layer wolf pack; The wolf pack obeys Layer wolf pack and The command of the wolf pack, responsible for reconnaissance and sentry; Surrounded by wolves Layer wolf pack, Layer wolf pack and Layer wolf packs and update their positions.

[0101] The hunting process of gray wolves includes surrounding prey, chasing prey and attacking prey.

[0102] (1) Surrounding the prey

[0103] During the hunting process, , and The wolves surround their prey according to Formula 3 and Formula 4.

[0104] (3)

[0105] (4)

[0106] Where: t is the number of iterations; D is the distance between the individual gray wolf and the prey; is the position vector of the prey; X is the position vector of the gray wolf; A is the first vector coefficient, C is the second vector coefficient, , ,in is the convergence factor, which decreases linearly from 2 to 0 as the number of iterations increases. is the first random vector, is the second random vector, modulo a random number between 0 and 1.

[0107] (2) Chasing prey

[0108] After the wolf is able to identify the location of its prey, , and The wolf pack surrounded the prey, and the individual gray wolf tracked the location of the prey according to formula 5 as follows.

[0109] (5)

[0110] Where: , and for , and the distance between the wolf and other individuals; , and for , and The wolf's current location; is the third random vector, is the fourth random vector, is the fourth random vector, Indicates the current location of the gray wolf. Wolf , and The compensation and direction of the wolf's advance are shown in Formula 6, The final position of the wolf is shown in Equation 7.

[0111] (6)

[0112] (7)

[0113] in, for The wolf's forward compensation vector, for The wolf's forward vector coefficient;

[0114] for The wolf's forward compensation vector, for The wolf's forward vector coefficient; for The wolf's forward compensation vector, for The wolf's forward vector coefficient; for The wolf's final position vector.

[0115] (3) Attacking prey

[0116] When the prey stops moving, the wolf completes the hunt by attacking. In the iterative process, As the number of iterations increases, A decreases linearly from 2 to 0. arrive Changes between. The larger the value of will make the gray wolf farther away from the prey. When , the gray wolf explores other areas to find the global optimal solution; if When the wolf pack attacks its prey.

[0117] Regarding the improvement of the Gray Wolf Algorithm, the chaotic population initialization method can generate initial solutions with randomness and complexity, which helps to increase the diversity of the population. Introducing the method of chaotic population initialization in the Gray Wolf Algorithm can improve the global search ability of the algorithm

[12] , so that the algorithm can obtain a wider search space in the initial stage, and then find a better solution, thereby enhancing the robustness of the Gray Wolf Algorithm. This application uses a sine mapping function to generate chaotic behavior, and its formula is shown in Formula 8.

[0118] (8)

[0119] Where: For the i The value of a state; is the system parameter, , in this application Take 0.99.

[0120] Nonlinear adjustment convergence factor, convergence factor in the Grey Wolf Algorithm It decreases linearly from 2 to 0 as the number of iterations increases, but the linear adjustment convergence factor cannot usually be flexibly adjusted according to the actual situation of the algorithm and the optimization process. The method of introducing a nonlinear adjustment convergence factor can enhance the adaptability and convergence speed of the algorithm, allowing the algorithm to better adapt to the search space and more accurately adjust the search strategy. This application adopts a nonlinear adjustment convergence factor based on a logarithmic function, and its calculation formula is shown in Formula 9.

[0121] (9)

[0122] Where: t is the number of current iterations; is the maximum number of iterations; and for The initial and final values ​​of are 2 and 0 respectively in this application.

[0123] Modify the position equation. The step length Euclidean distance principle is a dynamic adjustment strategy for the optimization algorithm. According to the parameter update in the optimization process, the proportional weight of the Euclidean distance is adaptively adjusted. In order to improve the robustness and convergence ability of the gray wolf algorithm, and more effectively explore the search space and find the optimal solution, the proportional weight dynamic adjustment strategy of the step length Euclidean distance principle is used to dynamically adjust the weight.

[0124] Optimize BP neural network. This application uses the Gray Wolf Algorithm to optimize the BP neural network as the basis for the thermal load prediction of the primary side of the thermal power station. The Gray Wolf Algorithm is improved through a multi-strategy fusion method to improve the prediction accuracy and obtain the optimal prediction value. The specific steps are as follows:

[0125] 1) Initialize the BP neural network topology. After determining that the gray wolf population is 150, the weights and thresholds of each layer of the BP neural network are used as the solution object of the gray wolf algorithm.

[0126] 2) Data preprocessing: Analyze and preprocess the acquired data, such as defect processing, outlier processing, normalization, etc.

[0127] 3) Initialize the wolf pack. The chaotic population initialization method is used in the improved grey wolf algorithm.

[0128] 4) Calculate the individual fitness of gray wolves according to the formula.

[0129] 5) Update the position of the gray wolf and use it as the input parameter of the BP neural network to train the network and obtain the updated individual fitness of the gray wolf.

[0130] 6) Determine whether the maximum number of iterations, 100, has been reached. If not, return to step 5. Otherwise, use the obtained optimal weights and thresholds as the parameters of the BP neural network to build the BP neural network.

[0131] 7) Train the Grey Wolf Algorithm to optimize the BP neural network model.

[0132] 8) Determine whether the model training structure meets the accuracy requirements. If not, return to retrain the model. Otherwise, output the results and end the training.

[0133] Evaluation indicators are metrics used to measure and evaluate the performance of models, algorithms, or systems. In different fields and tasks, a variety of evaluation indicators can be used. This application selects MAE, MSE, and coefficient of determination. The mean absolute percentage error (MAPE) is used as evaluation indicators to indicate the accuracy of the model prediction results. Its expressions are shown in formulas 10, 11, 12 and 13.

[0134] (10)

[0135] (11)

[0136] (12)

[0137] (13)

[0138] Where: is the actual value of heat load; is the predicted value of model heat load; is the average heat load; n is the number of samples.

[0139] This application adopts a method based on the improved Grey Wolf algorithm to optimize the BP neural network to predict the heat load on the primary side of the thermal power station. The thermal power station data is used as an example for verification, and the following conclusions are obtained: 1) The linear relationship between each eigenvector and the heat load is analyzed by the Pearson correlation coefficient method, and the eigenvectors that are strongly correlated with the heat load are used to predict future changes in the heat load. 2) The Grey Wolf algorithm is improved by a multi-strategy fusion method. The evaluation indicators of the IGWO-BP model are better than those of the GWO-BP model. The data proves that the improved algorithm solves the problems of being prone to falling into local optimality and being easily affected by linear convergence factors. At the same time, it is verified that it is effective and feasible in improving the accuracy of prediction. In summary, the method for predicting the heat load on the primary side of a thermal power station based on the improved Grey Wolf algorithm to optimize the BP neural network can improve the accuracy of the prediction and provide effective support for the optimized operation and energy-saving management of the thermal power station.

[0140] By executing the above steps, the data preprocessing link is adopted. Through time alignment, missing value interpolation, outlier processing, data smoothing and normalization, the quality and consistency of the original data are effectively improved, laying a good data foundation for subsequent modeling. Secondly, feature engineering technology is introduced. By using methods such as Pearson correlation coefficient analysis, decision tree and recursive feature elimination algorithm, the feature subset most relevant to the heat load is selected from many influencing factors as model input, avoiding the noise caused by redundant features, improving the representativeness and discrimination of features, and helping to improve the fitting and prediction capabilities of the model. In addition, the improved gray wolf optimization algorithm is organically combined with the BP neural network to play a complementary role in the advantages of the two. On the one hand, the gray wolf optimization algorithm has the characteristics of strong global optimization ability and fast convergence speed. By integrating various improvement measures such as chaos strategy population initialization, nonlinear adjustment convergence factor strategy and proportional weight dynamic adjustment strategy, it does not fall into the local optimal solution, and improves the robustness and adaptability of the algorithm. On the other hand, the BP neural network, with its powerful nonlinear mapping ability, can effectively mine the complex internal laws and dependencies in the heat load data. The combination of the two not only ensures that the algorithm has good global search capabilities, but also makes the model have stronger fitting capabilities, thereby improving prediction accuracy and stability.

[0141] At the same time, in the real-time prediction stage, the real-time operation data is processed and analyzed through technical means such as Kalman filtering, curve fitting, and feature extraction, which can not only capture the changing trend of heat load in a timely manner, but also identify mutation points and abnormal situations, so as to achieve continuous and accurate prediction and provide strong data support for heating scheduling decisions. Finally, an intuitive heat load forecast report is generated to facilitate heating management personnel to understand the changes in heat load, formulate corresponding scheduling strategies, and achieve accurate matching of supply and demand and optimal resource allocation.

[0142] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0143] (1) Collect data on the heating system of the target thermal power station to obtain historical operation data;

[0144] (2) Perform time alignment and frequency unification processing on the historical operation data to obtain aligned historical operation data;

[0145] (3) Interpolate missing values ​​of aligned historical operation data to obtain non-missing historical operation data;

[0146] (4) Perform outlier detection on the non-missing historical operation data to obtain the historical operation data with abnormal marks;

[0147] (5) Remove outliers from historical operation data with abnormal marks to obtain historical operation data without abnormalities;

[0148] (6) Performing data smoothing processing on the historical operation data without abnormalities according to the preset data smoothing index to obtain smoothed historical operation data;

[0149] (7) Segment the smoothed historical operation data to obtain training historical operation data and test historical operation data;

[0150] (8) Based on the preset normalization interval, the training history running data and the test history running data are respectively normalized to obtain a normalized training data set and a normalized test data set;

[0151] (9) Merge the normalized training data set and the normalized test data set to obtain a normalized historical operation data set;

[0152] (10) Perform data de-redundancy processing on the normalized historical operation data set to obtain a processed data set.

[0153] Specifically, the heating system of the target heating station is comprehensively collected to obtain historical operation data. Taking a heating company as an example, it is necessary to use data collection terminals and sensors to record the core operation parameters of multiple heating stations of the company in real time, such as water supply temperature, return water temperature, primary network water supply flow, heat metering value, etc., and summarize them into the original historical operation data set.

[0154] Since the data collection frequency and time of different thermal power stations may be different, it is necessary to time align and unify the original historical operation data. For example, if the water supply temperature sampling frequency of station A of a thermal power company is 1 minute, and that of station B is 5 minutes, then it is necessary to unify the two to the same time scale and sampling frequency through interpolation or downsampling to obtain the aligned historical operation data.

[0155] Next, due to communication failures, sensor failures, and other reasons, there may be some missing values ​​in the aligned historical operation data. In order to preserve the integrity and continuity of the data as much as possible, the scheme uses the method of calculating the average value of adjacent data to interpolate missing values. Taking the above-mentioned station A as an example, if the water supply temperature data is missing due to communication interruption at a certain point in time, the average value of the adjacent data before and after that time point can be used to estimate and fill in the missing values ​​to obtain historical operation data without missing values.

[0156] However, due to unexpected factors such as extreme weather and equipment failure, there may be outliers in the complete historical operation data. The outlier detection method based on the 3σ principle will be used to identify it, and the data that deviates from the normal value by more than 3 standard deviations will be marked as outliers, thus obtaining the historical operation data with anomaly marks.

[0157] For data with abnormal markers, outlier removal will be performed to obtain high-quality historical operation data without abnormalities. Taking Station B as an example, if an abnormal value of water supply temperature occurs due to extreme low temperatures in the cold winter, the abnormal data point will be automatically detected and removed.

[0158] In order to further improve data quality and availability, data smoothing will be performed on the historical operation data without abnormalities according to preset data smoothing indicators, such as moving average, to eliminate the influence of high-frequency noise and obtain smooth historical operation data. For example, the 5-point moving average smoothing of the primary network water supply flow data of Station A can effectively suppress the high-frequency noise caused by the pressure fluctuation of the pipe network.

[0159] The smoothed historical running data will be further divided into training data set and test data set. Taking the above data as an example, 75% of the data in the earlier period is usually used as the training set, and the latter 25% of the data is used as the test set for model training and performance evaluation.

[0160] Since the data input into the neural network needs to be normalized to eliminate the dimension effect, the training history running data and the test history running data will be normalized based on the preset normalization interval, such as [0, 1], to obtain the normalized training data set and test data set.

[0161] In order to obtain a complete historical operation data set, the normalized training data set and test data set are merged to form a normalized historical operation data set. Finally, the solution will perform data deduplication processing on the data set to remove redundant and noisy data and obtain the final processed data set, preparing for subsequent feature engineering and model building.

[0162] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0163] (1) Extract features from the processed data set to obtain the original feature set;

[0164] (2) Perform correlation analysis on the original feature set, calculate the Pearson correlation coefficient between each original feature and the heat load, and obtain the feature correlation matrix;

[0165] (3) Extract the diagonal elements of the feature correlation matrix to obtain the autocorrelation coefficient vector of each original feature;

[0166] (4) The autocorrelation coefficient vectors of each original feature are screened according to a preset threshold, and features with a value less than the preset threshold are removed to obtain a preliminary screening feature set;

[0167] (5) Construct a decision tree model based on the initial screening feature set, and analyze the decision tree model by processing the data set to obtain the information gain ratio of each original feature;

[0168] (6) Sort the initial screening feature set according to the information gain ratio of each original feature to obtain a list of ranked feature importance;

[0169] (7) Select the first N features in the feature importance list as the candidate feature set, where N is the preset feature number threshold;

[0170] (8) Recursively screen the candidate feature set through the recursive feature elimination algorithm until the preset number of iterations is met to obtain the optimal feature subset;

[0171] (9) Merge similar features of the selected feature subsets to obtain the target feature set.

[0172] Specifically, the most representative and discriminative target feature set is selected from the original feature set as the input of the prediction model. First, feature extraction is performed on the processed data set to obtain the original feature set. Taking a heating company as an example, the original feature set may include multiple features such as primary side temperature, primary side return temperature, primary side supply pressure, primary side return pressure, primary side valve position feedback, primary side instantaneous flow, heat load in the previous two hours, and heat load in the previous hour.

[0173] Since different features have different influences on the heat load, feature importance analysis is required. The scheme uses the Pearson correlation coefficient as the analysis indicator to calculate the linear correlation between each original feature and the actual heat load to obtain the feature correlation matrix. The diagonal elements of this matrix are the autocorrelation coefficients of each feature. These diagonal elements will be extracted to form the autocorrelation coefficient vectors of each original feature.

[0174] Next, the autocorrelation coefficient vector of each feature will be screened according to a pre-set correlation threshold, such as 0.6. Features with autocorrelation coefficients less than the threshold will be eliminated to obtain a preliminary screening feature set. Taking the above-mentioned heating company as an example, if the autocorrelation coefficients of the four features of primary side heating, primary side temperature recovery, heat load in the first two hours, and heat load in the first hour are all greater than 0.6, and other features are lower than the threshold, then these four features will enter the preliminary screening feature set, and the remaining features will be eliminated.

[0175] In order to further refine the feature set, a decision tree model is constructed based on the initial screening feature set, and the model is trained and analyzed using the processed data set to obtain the information gain ratio of each original feature in the decision tree. The information gain ratio can reflect the importance of each feature to the target variable (heat load). The higher the gain ratio, the more important the feature is.

[0176] According to the information gain ratio of each feature, the initial screening feature set is sorted to form a feature importance list. The first N features in the list are selected as the candidate feature set, where N is a preset feature number threshold, such as 10. Taking the above-mentioned heating company as an example, assuming that the initial screening feature set contains 8 features, after sorting according to the information gain ratio, the first 10 features with the highest importance will be selected to form the candidate feature set.

[0177] However, even after the above screening, there may still be redundant or highly correlated features in the candidate feature set. To further optimize the feature set, the recursive feature elimination algorithm will be used to recursively screen the candidate feature set until the preset number of iterations, such as 10 iterations, is met. The algorithm will gradually eliminate the features with the smallest contribution based on the cross-validation score of the model, and finally obtain the optimal feature subset.

[0178] After obtaining the optimal feature subset, the solution will also merge similar features. Since some features, although of equal importance, actually represent similar dimensions or concepts, they can be merged to avoid redundant information interfering with the model. Taking the above-mentioned heating company as an example, if the optimal subset contains both "primary side heating" and "primary side temperature return", since both actually reflect the temperature level of the primary side heating network, they can be merged into a "temperature difference" feature. Through this step, the solution finally obtains a streamlined and representative target feature set.

[0179] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0180] (1) Matching the node parameter set of the initial BP network according to the target feature set, where the node parameter set includes: the number of input nodes, the number of output nodes, and the number of hidden layer nodes;

[0181] (2) Initializing the parameters of the initial BP network through the node parameter set to obtain an initialization parameter set, wherein the initialization parameter set includes: network connection weights and network threshold data;

[0182] (3) By improving the gray wolf algorithm, the initialization parameter set is converted and encoded into gray wolf individuals to obtain the initial gray wolf population;

[0183] (4) Optimizing the initial gray wolf population through a preset chaotic mapping strategy to obtain an optimized gray wolf population;

[0184] (5) Extracting the fitness value of each optimized gray wolf individual in the optimized gray wolf population to obtain the fitness value of each optimized gray wolf individual;

[0185] (6) Adjust the search step size of each optimized gray wolf individual in the optimized gray wolf population, and update the gray wolf individual position of each optimized gray wolf individual according to the fitness value of each optimized gray wolf individual to obtain the target gray wolf population;

[0186] (7) decoding each target gray wolf in the target gray wolf population to obtain a target network parameter set, wherein the target network parameter set includes: a target network connection weight and a target network threshold data;

[0187] (8) Optimize the weight threshold of the initial BP network according to the target network parameter set to obtain the target BP network.

[0188] Specifically, the node parameter set of the initial BP network is matched according to the number of features, including the number of input nodes, the number of output nodes, and the number of hidden layer nodes. Taking a heating company as an example, assuming that the target feature set contains 4 features, such as primary side heating, temperature recovery, heat load in the first two hours, and heat load in the first hour, the corresponding number of input nodes should be set to 4; since the prediction target in this example is a single heat load value, the number of output nodes is set to 1; as for the number of hidden layer nodes, it can be set according to the complexity of the actual problem and experience, such as 7 hidden layer nodes.

[0189] After the node parameter set is determined, the initial BP network will be initialized to obtain the initialization parameter set, which includes the connection weights between the layers of the network and the threshold data of each node. These weights and thresholds are usually randomly initialized.

[0190] Next, we will optimize the above initialization parameter set based on the improved gray wolf optimization algorithm. First, we need to convert the initialization parameter set into a gray wolf individual encoding form suitable for the gray wolf algorithm operation, and initialize an initial gray wolf population containing multiple gray wolf individuals. The population size can be set to 150 gray wolf individuals.

[0191] In order to improve the global search capability of the algorithm, a chaos mapping strategy will be used to optimize the initial gray wolf population to produce an optimized gray wolf population with a more uniform distribution and greater diversity. Chaotic mapping can generate new solutions with randomness and chaotic characteristics through specific nonlinear equations, which helps the population to jump out of the local optimal area.

[0192] For each individual gray wolf in the optimized gray wolf population, its fitness value will be extracted, that is, the gray wolf individual is decoded and restored to the corresponding BP network parameters, and substituted into the set fitness function (such as the mean square error function) to calculate the fitness value, which is used to evaluate the quality of the individual. The smaller the fitness value, the better the individual.

[0193] Then, the search step size of each optimized gray wolf individual will be dynamically adjusted. The search step size determines the movement amplitude of the gray wolf individual in the solution space. If it is too large, it will fail to converge, and if it is too small, it will fall into the local optimum. Therefore, the current number of iterations, fitness value and other factors will be comprehensively considered to adaptively adjust the step size of each individual, so that the entire population can achieve a dynamic balance between global and local search.

[0194] According to the adjusted step length, combined with the fitness value of each optimized gray wolf individual, the position of each individual in the solution space will be updated to simulate the hunting behavior of the wolf pack and gradually converge to the optimal solution. This process will continue for several generations until the termination condition is met and the final target gray wolf population is obtained.

[0195] For each target gray wolf individual in the target gray wolf population, a decoding operation will be performed to restore it to the corresponding BP network parameter set, including the target network connection weight and target network threshold data. Taking the above-mentioned heating company as an example, a target gray wolf individual may correspond to a set of specific connection weight matrices and threshold vectors after decoding, which are used to describe a specific BP network structure.

[0196] Finally, the above-mentioned optimized target network parameter set will be used to adjust the weights and thresholds of the initial BP network structure to obtain a target BP network with better performance, preparing for subsequent model training and prediction.

[0197] This optimization process organically combines the improved Gray Wolf Algorithm with the BP neural network, giving full play to the complementary advantages of the two. On the one hand, the Gray Wolf Algorithm can efficiently find the optimal solution in the high-dimensional solution space with the help of population intelligence optimization technology; on the other hand, the BP neural network can effectively mine the inherent laws in complex data with its powerful nonlinear mapping ability. The fusion of the two not only improves the optimization performance of the algorithm, but also gives the model stronger learning and fitting capabilities, laying the foundation for accurate heat load prediction.

[0198] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0199] (1) Initializing the training parameters of the target BP network to obtain a training parameter set, where the training parameter set includes: learning rate, momentum factor and regularization coefficient;

[0200] (2) Based on the training parameter set, the normalized training data set is input into the target BP network for forward propagation to obtain the predicted output data;

[0201] (3) The loss gradient of the normalized training data set and the predicted output data is calculated using the preset loss function to obtain the gradient information;

[0202] (4) Based on the gradient information, the target BP network is updated through the back propagation algorithm to obtain an updated BP network;

[0203] (5) Perform model generalization processing on the updated BP network to obtain the initial heat load prediction model;

[0204] (6) Perform hyperparameter optimization on the initial heat load prediction model to obtain the heat load prediction model.

[0205] Specifically, after obtaining the target BP network structure optimized by the improved gray wolf algorithm, it will be trained to fully tap the mapping and fitting capabilities of the network. First, a series of training parameters of the target BP network need to be initialized to obtain a training parameter set. These parameters usually include learning rate, momentum factor, and regularization coefficient.

[0206] Among them, the learning rate determines the extent to which the weight is modified at each iteration. If it is too large, it will not converge, and if it is too small, the convergence process will be extremely slow. The momentum factor is used to smooth the gradient descent process and avoid falling into the local minimum to a certain extent. The regularization coefficient is a penalty term introduced to prevent overfitting.

[0207] Taking a heating company as an example, the learning rate can be initialized to 0.01, the momentum factor to 0.9, and the regularization coefficient to 0.001.

[0208] After the training parameter set is determined, the normalized training data set is input into the target BP network, and the forward propagation calculation is performed based on the training parameter set to obtain the corresponding predicted output data. Taking the above-mentioned heating company as an example, the training set part of the preprocessed historical operation data can be input into the BP network optimized by the gray wolf algorithm to obtain the corresponding heat load prediction value sequence.

[0209] Since neural networks are essentially supervised learning models, it is necessary to compare the predicted output data with the actual training label data. The loss function set in advance is used to calculate the loss of the normalized training data set (real label) and the predicted output data to obtain the gradient information of the prediction error. The loss function can be set to mean square error, cross entropy, etc., which is used to quantify the difference between the predicted value and the actual value.

[0210] With the gradient information, the connection weights and threshold parameters of the BP network will be adjusted along the gradient direction according to the principle of the back propagation algorithm, so that the difference between the output value and the true value is continuously reduced, thereby achieving the purpose of reducing losses and optimizing the network. This process uses the chain derivation rule to calculate the error responsibility of each node layer by layer from back to front, and updates the corresponding weights and threshold parameters. It is the core step of neural network training. By continuously iterating this process, the BP network will gradually fit the training data and obtain an updated BP network structure.

[0211] In order to avoid overfitting and improve the generalization ability of the network, the updated BP network will be subjected to generalization processing such as model regularization to enhance its ability to cope with new samples and obtain the initial heat load prediction model.

[0212] After obtaining the initial prediction model, the solution will also perform hyperparameter optimization on it to further improve the prediction accuracy and stability of the model. Hyperparameters refer to parameters that affect the training process outside the model training process, such as the number of training rounds, batch size, etc. Through optimization algorithms such as grid search and random search, various hyperparameter combinations will be traversed within a certain range to evaluate their impact on model performance, and the optimal hyperparameter configuration will be selected to finally obtain a heat load prediction model with good performance.

[0213] The training process is consistent with the basic concepts of machine learning, such as forward propagation, loss function calculation, back propagation, and parameter update, and incorporates the latest technologies in the fields of regularization and hyperparameter optimization to ensure the model's fitting effect and generalization performance. At the same time, the process reasonably sets key training parameters such as learning rate and loss function, and illustrates the specific values ​​and meanings of these parameters in practical problems through examples, improving operability and practicality.

[0214] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0215] (1) Real-time collection of current operating data of the target thermal power station, including: water supply temperature, return water temperature, primary network water supply flow and heat;

[0216] (2) Perform data stream conversion on the current running data to obtain the real-time running data stream;

[0217] (3) Filtering the real-time operation data stream through the Kalman filter algorithm to obtain a smooth real-time operation data stream;

[0218] (4) extracting data segments from the smooth real-time running data stream to obtain real-time running segment data;

[0219] (5) Extracting operation features from the real-time operation segment data to obtain an operation feature vector set;

[0220] (6) Input the operating feature vector set into the heat load prediction model for forward propagation to obtain a set of heat load prediction values;

[0221] (7) Perform curve fitting on the heat load prediction value set to obtain the heat load prediction curve.

[0222] Specifically, the current operating data of the target thermal power station is collected in real time. These data usually include core operating parameters such as water supply temperature, return water temperature, primary network water supply flow and heat metering value. Taking a heating company as an example, the above four key data can be obtained in real time through the online detection system deployed in each thermal power station.

[0223] Since the raw data collected in real time may be interfered by noise, in order to improve the data quality, the current running data will be converted into a real-time running data stream. Common methods of data stream conversion include moving average method, exponential weighted moving average method, etc., which can perform preliminary smoothing on continuous time series data.

[0224] In order to further improve the data quality and reliability, the Kalman filter algorithm will be introduced to optimize the filtering of the real-time operation data stream to obtain a smooth real-time operation data stream. Kalman filtering is a commonly used optimal estimation algorithm. By minimizing the feedback of the prediction error, it can effectively remove the noise in the data and reconstruct a reliable signal. Taking the above-mentioned heating company as an example, the Kalman filter algorithm can be applied to the real-time collected supply and return water temperature data stream to eliminate the high-frequency noise introduced by the temperature sensor error and restore the real and reliable temperature curve.

[0225] After obtaining a smooth real-time running data stream, appropriate data segments will be extracted from it to obtain real-time running segment data. For example, if the prediction model input features include "heat load in the last two hours" and "heat load in the last hour", it is necessary to extract the data segments of the last few hours from the smooth data stream.

[0226] Next, the operation feature extraction is performed on the real-time operation segment data to obtain an operation feature vector set that matches the prediction model input. Taking a certain thermal power station as an example, if the input features of the prediction model are heating, temperature recovery, heat load in the previous two hours, and heat load in the previous hour, the current values ​​of the above four features are extracted from the real-time data segment respectively and spliced ​​to form an operation feature vector.

[0227] After obtaining the operating feature vector set, it is input into the trained heat load prediction model to perform forward propagation calculation, thereby obtaining a set of corresponding heat load prediction value sets, which represent the predicted heat load level in the future period of time.

[0228] In order to obtain a more continuous and smooth heat load change trend, the above predicted value set will be subjected to curve fitting to obtain the final heat load prediction curve. Commonly used curve fitting methods include polynomial fitting and cubic spline interpolation, which can fully capture the change law of the predicted value and eliminate mutations and fluctuations.

[0229] Taking the above-mentioned heating company as an example, assuming that the prediction model outputs the predicted heat load value for each hour in the next 24 hours, the solution can fit these 24 prediction points into a smooth curve, clearly reflecting the overall picture and changing trend of the heat load in the next day, and providing intuitive data support for thermal scheduling decisions.

[0230] The real-time prediction link closely combines the latest real-time operation data with the trained prediction model to form a closed-loop online prediction system. Through the careful design of data stream processing, feature extraction, model forward calculation and curve fitting, the real-time, continuity and reliability of the prediction results are ensured.

[0231] At the same time, this link also integrates a variety of digital signal processing and time series analysis techniques, such as Kalman filtering and moving average, which effectively improves the quality of real-time data and provides high-quality input for model prediction.

[0232] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0233] (1) Perform discrete sampling of the predicted heat load curve to obtain discrete prediction data;

[0234] (2) Perform piecewise linear regression fitting on the discretized prediction data to generate a local trend feature vector;

[0235] (3) Extract statistical feature vectors from local trend feature vectors to obtain global statistical feature vectors;

[0236] (4) Perform fast Fourier transform on the global statistical eigenvector to obtain the periodic eigenvector;

[0237] (5) Cluster the periodic feature vector into subsequences of equal length to obtain the pattern feature vector;

[0238] (6) Generate a heat load prediction report based on the pattern feature vector, and transmit the heat load prediction report to a preset data display terminal.

[0239] It should be noted that a series of data processing and feature extraction operations will be performed on the curve. First, the predicted heat load curve will be sampled discretely to obtain a series of discrete prediction data points. Taking a heating company as an example, assuming that the prediction curve reflects the predicted heat load value per hour in the next 24 hours, it can be discretized into 24 data points.

[0240] After the discretized forecast data is obtained, it will be fitted with a piecewise linear regression to generate a local trend feature vector. Linear regression can effectively capture the local linear trend of the data and better characterize the nonlinear change characteristics of the data through piecewise fitting. Taking the above-mentioned heating company as an example, the 24-hour forecast data can be piecewise linearly regressed every 6 hours to obtain the parameters such as the slope and intercept of the 4 local trend line segments, forming an 8-dimensional local trend feature vector.

[0241] Next, we will extract statistical features from the local trend feature vector to obtain a global statistical feature vector. Statistical features can reflect the global distribution characteristics of the data, such as mean, standard deviation, skewness, kurtosis, etc., which helps to quantify the macro-change trend of the data. Continuing with the heating company as an example, we can extract statistical quantities such as mean, maximum value, minimum value from the local trend vector and splice them to form a 5-dimensional global statistical feature vector.

[0242] In order to further explore the inherent periodic characteristics of the forecast data, the global statistical feature vector is subjected to fast Fourier transform (FFT) to obtain the periodic feature vector. FFT can decompose the time domain signal into the amplitude and phase information of different frequency components, thereby revealing the potential periodic laws in the data. Taking the above-mentioned heating company as an example, the 5-dimensional statistical vector can be subjected to FFT transformation to extract the frequency components corresponding to the first three largest amplitudes to form a 3-dimensional periodic feature vector.

[0243] Finally, the periodic feature vectors are clustered into equal-length subsequences to obtain pattern feature vectors. Equal-length subsequence clustering can be used to discover typical patterns in data and provide support for classification and anomaly detection of time series data. Taking the above-mentioned heating company as an example, the 3D periodic vector can be divided into subsequences of length 2 and clustered based on distance metrics to obtain cluster centroids representing different periodic patterns, that is, 2D pattern feature vectors.

[0244] Based on the results of the above series of feature extraction and analysis, an intuitive heat load forecast report will be generated and transmitted to the preset data display terminal to provide decision support for heating management personnel. The report can include a variety of visual display forms such as heat load forecast curves, local trend segments, statistical feature radar charts, and periodic pattern legends, which fully reflect the distribution characteristics and change laws of the forecast data and provide a comprehensive data basis for formulating scheduling strategies.

[0245] This link combines a variety of data processing and mining technologies such as discrete sampling, linear regression, statistical feature extraction, FFT analysis and clustering to conduct a multi-dimensional and multi-level analysis of the forecast curve, quantitatively characterizing the characteristics of the forecast data from different angles, which can be regarded as an in-depth analysis and sublimation of the forecast results. Compared with the original forecast curve, the forecast report output by this link is not only more intuitive in form and richer in content, but also reflects the statistical characteristics, time series characteristics and pattern characteristics of the forecast data, and has higher theoretical value and application guidance significance.

[0246] The embodiment of the present invention further provides a thermal power station primary side heat load prediction system, and the thermal power station primary side heat load prediction system specifically includes:

[0247] The acquisition module 201 is used to acquire the historical operation data of the target thermal power station and pre-process the historical operation data to obtain a processed data set;

[0248] A screening module 202 is used to perform feature importance analysis and screening on the processed data set to obtain a target feature set;

[0249] A construction module 203 is used to construct an initial BP network according to the target feature set, and optimize the weight threshold of the initial BP network by a preset improved grey wolf algorithm to obtain a target BP network;

[0250] A training module 204 is used to perform model training on the target BP network through the processed data set to obtain a heat load prediction model;

[0251] The prediction module 205 is used to collect the current operation data of the target thermal power station in real time, and input the current operation data into the thermal load prediction model to perform thermal load prediction, so as to obtain a thermal load prediction curve;

[0252] The extraction module 206 is used to extract heat load prediction data according to the heat load prediction curve, generate a heat load prediction report, and transmit the heat load prediction report to a preset data display terminal.

[0253] Through the collaborative work of the above modules, the data preprocessing link is adopted. Through time alignment, missing value interpolation, outlier processing, data smoothing and normalization, the quality and consistency of the original data are effectively improved, laying a good data foundation for subsequent modeling. Secondly, feature engineering technology is introduced. By using methods such as Pearson correlation coefficient analysis, decision tree and recursive feature elimination algorithm, the feature subset most relevant to the heat load is selected from many influencing factors as model input, avoiding the noise caused by redundant features, improving the representativeness and distinguishability of features, and helping to improve the fitting and prediction capabilities of the model. In addition, the improved gray wolf optimization algorithm is organically combined with the BP neural network to play a complementary role in the advantages of the two. On the one hand, the gray wolf optimization algorithm has the characteristics of strong global optimization ability and fast convergence speed. By integrating various improvement measures such as chaos strategy population initialization, nonlinear adjustment convergence factor strategy and proportional weight dynamic adjustment strategy, it does not fall into the local optimal solution, and improves the robustness and adaptability of the algorithm. On the other hand, the BP neural network, with its powerful nonlinear mapping ability, can effectively mine the complex internal laws and dependencies in the heat load data. The combination of the two not only ensures that the algorithm has good global search capabilities, but also makes the model have stronger fitting capabilities, thereby improving prediction accuracy and stability.

[0254] At the same time, in the real-time prediction stage, the real-time operation data is processed and analyzed through technical means such as Kalman filtering, curve fitting, and feature extraction, which can not only capture the changing trend of heat load in a timely manner, but also identify mutation points and abnormal situations, so as to achieve continuous and accurate prediction and provide strong data support for heating scheduling decisions. Finally, an intuitive heat load forecast report is generated to facilitate heating management personnel to understand the changes in heat load, formulate corresponding scheduling strategies, and achieve accurate matching of supply and demand and optimal resource allocation.

[0255] The above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the embodiments, a person skilled in the art should understand that the specific implementation modes of the present invention can still be modified or replaced by equivalents, and any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A method for predicting heat load on the primary side of a thermal power station, characterized in that: include: Collecting historical operation data of the target thermal power station, and preprocessing the historical operation data to obtain a processed data set; The processing data set is subjected to feature importance analysis and screening to obtain a target feature set, including: extracting features from the processing data set to obtain an original feature set; performing correlation analysis on the original feature set, calculating the Pearson correlation coefficient between each original feature and the heat load, and obtaining a feature correlation matrix; extracting diagonal elements from the feature correlation matrix to obtain an autocorrelation coefficient vector of each original feature; screening the autocorrelation coefficient vector of each original feature according to a preset threshold, removing features less than the preset threshold, and obtaining a preliminary screening feature set; constructing a decision tree model based on the preliminary screening feature set, and analyzing the decision tree model through the processing data set to obtain an information gain ratio of each original feature; sorting the preliminary screening feature set according to the information gain ratio of each original feature to obtain a sorted feature importance list; selecting the first N features in the feature importance list as a candidate feature set, wherein N is a preset feature quantity threshold; recursively screening the candidate feature set through a recursive feature elimination algorithm until a preset number of iterations is met to obtain a preferred feature subset; merging similar features of the preferred feature subset to obtain the target feature set; An initial BP network is constructed according to the target feature set, and a weight threshold of the initial BP network is optimized by a preset improved grey wolf algorithm to obtain a target BP network; Performing model training on the target BP network by using the processed data set to obtain a heat load prediction model; The current operation data of the target thermal power station is collected in real time, and the current operation data is input into the thermal load prediction model to perform thermal load prediction to obtain a thermal load prediction curve, including: collecting the current operation data of the target thermal power station in real time, wherein the current operation data includes: water supply temperature, return water temperature, primary network water supply flow and heat; performing data stream conversion on the current operation data to obtain a real-time operation data stream; filtering the real-time operation data stream through a Kalman filter algorithm to obtain a smooth real-time operation data stream; extracting data segments from the smooth real-time operation data stream to obtain real-time operation segment data; extracting operation features from the real-time operation segment data to obtain an operation feature vector set; inputting the operation feature vector set into the thermal load prediction model for forward propagation to obtain a thermal load prediction value set; performing curve fitting on the thermal load prediction value set to obtain the thermal load prediction curve; Extracting heat load prediction data according to the heat load prediction curve, generating a heat load prediction report, and transmitting the heat load prediction report to a preset data display terminal, including: performing dot matrix discrete sampling on the predicted heat load curve to obtain discrete prediction data; performing piecewise linear regression fitting on the discretized prediction data to generate a local trend feature vector; performing statistical feature vector extraction on the local trend feature vector to obtain a global statistical feature vector; performing fast Fourier transform on the global statistical feature vector to obtain a periodic feature vector; performing equal-length subsequence clustering on the periodic feature vector to obtain a pattern feature vector; generating the heat load prediction report according to the pattern feature vector, and transmitting the heat load prediction report to a preset data display terminal.

2. The method for predicting the primary side heat load of a thermal power station according to claim 1, characterized in that: The step of collecting the historical operation data of the target thermal power station and preprocessing the historical operation data to obtain a processed data set includes: Collecting data on the heating system of the target thermal power station to obtain historical operation data; performing time alignment and frequency unification processing on the historical operation data to obtain aligned historical operation data; Interpolating missing values ​​on the aligned historical operation data to obtain non-missing historical operation data; Performing outlier detection on the non-missing historical operation data to obtain historical operation data containing anomaly marks; Performing outlier elimination processing on the historical operation data containing abnormal marks to obtain historical operation data without abnormalities; Performing data smoothing processing on the non-abnormal historical operation data according to a preset data smoothing index to obtain smoothed historical operation data; Performing data segmentation on the smoothed historical operation data to obtain training historical operation data and test historical operation data; Based on a preset normalization interval, the training historical operation data and the test historical operation data are respectively subjected to data normalization processing to obtain a normalized training data set and a normalized test data set; Merging the normalized training data set and the normalized test data set to obtain a normalized historical operation data set; The normalized historical operation data set is subjected to data de-redundancy processing to obtain a processed data set.

3. The method for predicting the primary side heat load of a thermal power station according to claim 1, characterized in that: The step of constructing an initial BP network according to the target feature set and optimizing the weight threshold of the initial BP network by a preset improved grey wolf algorithm to obtain a target BP network comprises: Matching the node parameter set of the initial BP network according to the target feature set, wherein the node parameter set includes: the number of input nodes, the number of output nodes and the number of hidden layer nodes; Initializing the parameters of the initial BP network through the node parameter set to obtain an initialization parameter set, wherein the initialization parameter set includes: network connection weights and network threshold data; The initialization parameter set is converted and encoded into gray wolf individuals by using the improved gray wolf algorithm to obtain an initial gray wolf population; The initial gray wolf population is optimized by a preset chaotic mapping strategy to obtain an optimized gray wolf population; Extracting the fitness value of each optimized gray wolf individual in the optimized gray wolf population to obtain the fitness value of each optimized gray wolf individual; Adjusting the search step length of each optimized gray wolf individual in the optimized gray wolf population, and updating the gray wolf individual position of each optimized gray wolf individual according to the fitness value of each optimized gray wolf individual, to obtain a target gray wolf population; Decoding each target gray wolf in the target gray wolf population to obtain a target network parameter set, wherein the target network parameter set includes: a target network connection weight and a target network threshold data; The initial BP network is optimized by weight threshold according to the target network parameter set to obtain a target BP network.

4. The method for predicting the primary side heat load of a thermal power station according to claim 2, characterized in that: The step of training the target BP network model by processing the data set to obtain a heat load prediction model includes: Initializing the training parameters of the target BP network to obtain a training parameter set, wherein the training parameter set includes: a learning rate, a momentum factor, and a regularization coefficient; Based on the training parameter set, the normalized training data set is input into the target BP network for forward propagation to obtain predicted output data; the loss gradient of the normalized training data set and the predicted output data is calculated by a preset loss function to obtain gradient information; According to the gradient information, the target BP network is updated by a back propagation algorithm to obtain an updated BP network; Performing model generalization processing on the updated BP network to obtain an initial heat load prediction model; The initial heat load prediction model is subjected to hyperparameter optimization processing to obtain the heat load prediction model.

5. A thermal power station primary side heat load prediction system, used to execute the thermal power station primary side heat load prediction method according to any one of claims 1 to 4, characterized in that: include: A collection module, used for collecting historical operation data of a target thermal power station, and preprocessing the historical operation data to obtain a processed data set; A screening module is used to perform feature importance analysis and screening on the processed data set to obtain a target feature set, including: extracting features from the processed data set to obtain an original feature set; performing correlation analysis on the original feature set to calculate the Pearson correlation coefficient between each original feature and the heat load to obtain a feature correlation matrix; extracting diagonal elements from the feature correlation matrix to obtain an autocorrelation coefficient vector of each original feature; screening the autocorrelation coefficient vector of each original feature according to a preset threshold, removing features less than the preset threshold, and obtaining a preliminary screening feature set; constructing a decision tree model based on the preliminary screening feature set, and analyzing the decision tree model through the processed data set to obtain an information gain ratio of each original feature; sorting the preliminary screening feature set according to the information gain ratio of each original feature to obtain a sorted feature importance list; selecting the first N features in the feature importance list as a candidate feature set, wherein N is a preset feature quantity threshold; recursively screening the candidate feature set through a recursive feature elimination algorithm until a preset number of iterations is met to obtain a preferred feature subset; merging similar features of the preferred feature subset to obtain the target feature set; A construction module is used to construct an initial BP network according to the target feature set, and optimize the weight threshold of the initial BP network through a preset improved grey wolf algorithm to obtain a target BP network; A training module, used for performing model training on the target BP network through the processed data set to obtain a heat load prediction model; The prediction module is used to collect the current operation data of the target thermal power station in real time, and input the current operation data into the thermal load prediction model to perform thermal load prediction to obtain a thermal load prediction curve, including: collecting the current operation data of the target thermal power station in real time, wherein the current operation data includes: water supply temperature, return water temperature, primary network water supply flow and heat; performing data stream conversion on the current operation data to obtain a real-time operation data stream; filtering the real-time operation data stream through a Kalman filter algorithm to obtain a smooth real-time operation data stream; extracting data segments from the smooth real-time operation data stream to obtain real-time operation segment data; extracting operation features from the real-time operation segment data to obtain an operation feature vector set; inputting the operation feature vector set into the thermal load prediction model for forward propagation to obtain a thermal load prediction value set; performing curve fitting on the thermal load prediction value set to obtain the thermal load prediction curve; The extraction module is used to extract heat load prediction data according to the heat load prediction curve, generate a heat load prediction report, and transmit the heat load prediction report to a preset data display terminal, including: performing dot matrix discrete sampling on the predicted heat load curve to obtain discrete prediction data; performing piecewise linear regression fitting on the discretized prediction data to generate a local trend feature vector; performing statistical feature vector extraction on the local trend feature vector to obtain a global statistical feature vector; performing fast Fourier transform on the global statistical feature vector to obtain a periodic feature vector; performing equal-length subsequence clustering on the periodic feature vector to obtain a pattern feature vector; generating the heat load prediction report according to the pattern feature vector, and transmitting the heat load prediction report to a preset data display terminal.

Images