A joint optimization method and system based on peak-shaving power equipment combination prediction

Abstract

The application discloses a kind of based on peak-shaving power equipment combination prediction joint optimization method and system, it is related to the state monitoring technical field of hydroelectric generating set.According to the different characteristics of hydroelectric generating set parameters, ARIMA model, random forest model, LSTM neural network model are established to carry out parameter prediction and weighted average, and the predicted parameters are obtained;Then the predicted parameters are input into the support vector machine SVM model and XGBoost model to classify and early warn the hydroelectric generating set fault, and the final fault classification early warning result is obtained by information fusion of the early warning result using the weighted evidence theory method.Through the calculation and fusion of multiple models, the problem of insufficient classification early warning accuracy and low reliability of single model classification is overcome, accurate and reliable parameter prediction and state early warning of peak-shaving power equipment, especially hydroelectric generating set equipment, are realized, which provides strong support for the safe and stable operation of power equipment.

Description

Technical Field

[0001] This invention belongs to the field of hydropower unit condition monitoring technology, specifically relating to a joint optimization method and system based on peak-shaving power equipment combination prediction. Background Technology

[0002] Hydropower units are among the key pieces of equipment in a power system, and their operational status directly impacts the reliability of the power grid's power supply. Therefore, employing reliable condition monitoring technologies for analysis and prediction of their future operational status, timely detection of faults, prevention of further escalation of accidents, and avoidance of major power incidents are crucial for ensuring the normal operation of hydropower units and for peak-shaving power supply in the power grid.

[0003] However, traditional hydropower unit diagnostic methods rely on theoretical knowledge of the equipment and human experience, resulting in poor universality and low efficiency, failing to effectively predict the state of hydropower units and provide timely early warnings. Data mining-based state monitoring methods mainly fall into three categories: statistical analysis methods, limited by statistical analysis theory, only achieve high accuracy on specific datasets and have poor generalization ability; signal processing methods characterize system state by extracting time-frequency domain feature parameters, but information loss easily occurs during time-frequency transformations; and machine learning-based equipment monitoring methods are numerous, but suffer from poor generalization of single prediction models and insufficient reliability. Therefore, the aforementioned manual or data mining-based detection methods cannot be widely and accurately used for predicting the operating state of hydropower units.

[0004] Therefore, how to propose a multi-model combination method for predicting the operating status of hydropower units, thereby improving the reliability and credibility of hydropower unit operating status prediction, is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a joint optimization method and system based on the combined prediction of peak-shaving power equipment, which uses data mining to predict the parameters and provide early warning of the status of peak-shaving power equipment, especially hydropower units, thereby improving the reliability and credibility of the prediction of the operating status of hydropower units.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention discloses a joint optimization method based on peak-shaving power equipment combination prediction, the specific steps of which are as follows:

[0008] Obtain the internal operating parameters, external environmental parameters, and power grid load parameters of the hydropower unit, and construct a data table;

[0009] Based on the data table, determine the initial dataset;

[0010] Based on the linear, nonlinear, and complex features of the data in the initial dataset, ARIMA, Random Forest, and LSTM neural network models are constructed to predict parameters; combined weights are calculated, and the predicted parameters are weighted and averaged to obtain the final output, which serves as the prediction dataset.

[0011] Multiple types of hydropower unit faults are defined as a fault space. Based on the initial dataset, support vector machine (SVM) and XGBoost models are built respectively, and the parameters of the two models are optimized using an optimization algorithm. According to the prediction dataset and fault space, the faults of hydropower units are classified and early warnings are issued, and initial early warning results are obtained respectively. The initial early warning results are fused to output the final state early warning results.

[0012] Furthermore, the internal operating parameters include: water turbine inlet pressure, vibration, and slewing coefficient data; generator rotor AC impedance and magnetic pole coefficient data; and transformer oil temperature and gas concentration data. The environmental parameters include: air temperature, ground temperature, relative humidity, and average wind speed data. The grid load parameters include: operating voltage, active power, and reactive power. Then, a data table is constructed based on the acquisition time of each parameter.

[0013] Furthermore, determining the initial dataset includes: first, determining the proportion of missing values ​​in the data table to the total sample size; when the proportion is greater than a threshold, using imputation to fill in the missing values ​​in the dataset; when it is less than the threshold, using deletion to directly delete the missing values; then, handling outliers; next, using the Z-score method to standardize the data; and finally, using out-of-bag feature permutation to extract features to obtain the initial dataset.

[0014] Furthermore, the construction of the ARIMA model, random forest model, and LSTM neural network model for parameter prediction includes:

[0015] Parameters with linear features are selected from the initial dataset, an ARIMA model is constructed and trained, and the trained ARIMA model is used for parameter prediction; parameters with nonlinear features are selected from the initial dataset, a random forest model is constructed and trained, and the trained random forest model is used for parameter prediction; parameters with complex features are selected from the initial dataset, an LSTM neural network model is constructed and trained, and the trained LSTM neural network model is used for parameter prediction.

[0016] The calculation of combined weights involves weighted averaging of the prediction parameters to obtain the final output, including:

[0017] The weights of the ARIMA model, random forest model, and LSTM neural network model are assigned using a goodness-of-fit algorithm. First, the coefficient of determination p of the goodness-of-fit algorithm is calculated. The normal range of p is [0,1]. The calculation formula is as follows:

[0018]

[0019] in, Predict the value for the i-th parameter. z is the average of the true values. i Let be the true value of the i-th parameter, and n be the total number of predicted parameters;

[0020] Then, the tangent function is used to optimize the coefficient of determination p, and finally the optimized formula for calculating the combined weights is obtained:

[0021]

[0022] Among them, w j z represents the combined weight of the j-th sub-model. j h(z) represents the p-value of the j-th sub-model. j ) represents the p-value of the j-th sub-model after transformation using the tangent function;

[0023] Based on the weight value w j The weighted average of the prediction parameters of the ARIMA model, the random forest model, and the LSTM neural network model is calculated to obtain the final output result, which is used as the prediction dataset.

[0024] Furthermore, the ARIMA model consists of an autoregressive model and a moving average model, with the initial coefficients including the number of autoregressive terms and the number of moving average terms; the random forest model uses bootstrap sampling to construct each basic decision tree; and the LSTM neural network model is optimized using the stochastic gradient descent Adam algorithm, with the loss function being the cross-entropy function.

[0025] Furthermore, the fault space is Θ{F1, F2, ..., F10} as the fault identification framework for hydropower units;

[0026] The Support Vector Machine (SVM) model is constructed by using a grid optimization method to roughly select the penalty parameter c and the kernel function parameter g; then, the K-fold cross-validation method is used to select the optimal c and g values ​​as the optimal parameters of the SVM model.

[0027] The XGBoost model is constructed by continuously updating the number of base learners, learning rate, and maximum tree depth using an adaptive particle swarm optimization algorithm to determine the optimal parameters.

[0028] The initial warning results include: using the optimized support vector machine (SVM) model and XGBoost model, with the final output of the prediction model of the prediction dataset parameters as input, to classify and warn of the faults of the hydropower unit, and obtain two initial warning results respectively;

[0029] The final state warning result is obtained by fusing information from two initial warning results using a weighted evidence method.

[0030] Furthermore, the K-fold cross-validation method specifically involves: dividing the data in the initial dataset into k subsets, where k-1 subsets are used as training subsets and 1 subset is used as test subsets; selecting a set of values ​​for c and g for training the support vector machine (SVM) model; calculating the mean squared error (MSE) of the test results of the test subset; calculating the MSE of multiple sets of different values ​​for c and g; and selecting the optimal c and g values ​​based on the principle of minimizing the MSE.

[0031] Furthermore, the adaptive particle swarm optimization algorithm is specifically as follows:

[0032] The inertia weight w for the adaptive particle swarm optimization algorithm is determined by the following formula:

[0033]

[0034] Among them, w max and w win These represent the maximum and minimum values ​​of the inertia weight, respectively, where k is the current iteration number. max This represents the maximum number of iterations.

[0035] Based on the evolution and aggregation degree of the particle swarm, the inertia weight of the adaptive particle swarm optimization algorithm is dynamically updated. The improved adaptive inertia weight update formula is as follows:

[0036] w * =w in -(w in -w min )*evol+(w max -w in )*aggr

[0037] Among them, w * w represents the adaptive inertia weight. in This represents the initial inertia weight value, where evol and aggr represent the degree of evolution and the degree of aggregation, respectively.

[0038] Furthermore, the weighted evidence method specifically includes:

[0039] In the two initial warning results, for each type of hydropower unit fault, there are two evidence bodies E1 and E2 to be combined, with corresponding basic probability allocation functions m1 and m2, and corresponding focal elements Ap and Bq, respectively. The common focal element of Ap and Bq is C, and the conflict coefficient between m1 and m2 is H. Then the fusion formula is:

[0040]

[0041] in, This represents the average level of support for the focal element from the evidence sources.

[0042] By using a fusion formula, early warning results for different types of hydropower unit faults are obtained through fusion.

[0043] Furthermore, it also includes evaluating the early warning performance of hydropower units by calculating average accuracy indicators, which include: average accuracy, precision, recall, and harmonic mean F1-score.

[0044] This invention also discloses a joint optimization system based on peak-shaving power equipment combination prediction, comprising:

[0045] Data acquisition module: Acquires internal operating parameters, external environmental parameters, and power grid load parameters of the hydropower unit, and constructs data tables;

[0046] Data preprocessing module: Determines the initial dataset based on the data table;

[0047] Parameter prediction module: Based on the linear, nonlinear and complex features of the data in the initial dataset, construct ARIMA model, random forest model and LSTM neural network model to predict parameters; calculate the combined weights, perform a weighted average of the predicted parameters, and obtain the final output result as the prediction dataset;

[0048] Status warning module: Defines various types of hydropower unit faults into a fault space, and builds support vector machine (SVM) and XGBoost models respectively based on the initial dataset. Optimization algorithms are used to optimize the parameters of the two models. Based on the prediction dataset and fault space, the module classifies and warns of hydropower unit faults, obtaining initial warning results. The initial warning results are then fused to output the final status warning result.

[0049] Furthermore, the parameter prediction module includes:

[0050] ARIMA model unit: Select parameters with linear features from the initial dataset, construct and train the ARIMA model, and use the trained ARIMA model to predict parameters.

[0051] Random forest model unit: Select parameters with nonlinear characteristics from the initial dataset, construct and train a random forest model, and use the trained random forest model to predict parameters;

[0052] LSTM neural network model unit: Select parameters with complex features from the initial dataset, construct and train an LSTM neural network model, and use the trained LSTM neural network model to predict parameters;

[0053] Combined weight unit: The weights of the ARIMA model, random forest model, and LSTM neural network model are assigned using a goodness-of-fit algorithm. Based on the weights, the weighted average of the prediction parameters of the ARIMA model, random forest model, and LSTM neural network model is calculated to obtain the final output result, which serves as the prediction dataset.

[0054] Furthermore, the status warning module includes:

[0055] SVM model unit: The grid optimization method is used to roughly select the penalty parameter c and the kernel function parameter g; then the K-fold cross-validation method is used to select the optimal c and g values ​​as the optimal parameters of the support vector machine SVM model. The prediction dataset is used as input to classify and warn of faults in hydropower units.

[0056] XGBoost model unit: The adaptive particle swarm optimization algorithm is used to continuously update the number of base learners, learning rate and maximum tree depth of the XGBoost model to determine the optimal parameters. The prediction dataset is used as input to classify and warn of faults in hydropower units.

[0057] Early warning result fusion unit: By using the weighted evidence method, the initial early warning results obtained from the SVM model unit and the XGBoost model unit are fused to obtain the final state early warning result of the hydropower unit.

[0058] As can be seen from the above technical solution, compared with the prior art, this invention discloses a joint optimization method and system based on the combined prediction of peak-shaving power equipment. By preprocessing and extracting features from historical parameter data of hydropower units, and classifying them according to linear, nonlinear, and complex features, different parameter prediction models are established for parameter prediction and weighted fusion, enabling more accurate prediction of future data for relevant parameters of hydropower units. By inputting the predicted parameters into a Support Vector Machine (SVM) model and an XGBoost model for hydropower unit fault early warning, and employing the weighted evidence theory to fuse the early warning results of the two models, the accuracy and reliability of hydropower unit early warning are further improved. The fusion of multiple models overcomes the problems of insufficient reliability of a single prediction model and low accuracy and reliability of classification early warning. Through data mining, parameter prediction and status early warning for peak-shaving power equipment, especially hydropower units, are achieved, providing strong support for the safe and stable operation of hydropower units and the power system. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0060] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0061] Figure 2 This is a schematic diagram of the status early warning process of the present invention. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] This invention discloses a joint optimization method based on peak-shaving power equipment combination prediction. The overall process is as follows: Figure 1 As shown, the specific steps are as follows:

[0064] Obtain the internal operating parameters, external environmental parameters, and power grid load parameters of the hydropower unit, and construct a data table.

[0065] The initial dataset is obtained by preprocessing and feature extraction of the data in the data table.

[0066] Construct parameter prediction models: Based on the linear, nonlinear, and complex features of the data in the initial dataset, construct ARIMA, Random Forest, and LSTM neural network models to predict parameters; calculate the combined weights, perform a weighted average of the predicted parameters, and obtain the final output results as the prediction dataset.

[0067] A state early warning model is constructed, defining various types of hydropower unit faults as a fault space. Based on the initial dataset, a support vector machine (SVM) model and an XGBoost model are built respectively, and the parameters of the two models are optimized using an optimization algorithm. According to the prediction dataset and the fault space, the faults of the hydropower units are classified and early warnings are issued, and initial early warning results are obtained respectively. The initial early warning results of the two models are fused to output the final state early warning result.

[0068] In one embodiment of the present invention, internal operating parameters of the hydropower unit are acquired, including: inlet water pressure, vibration, and slewing coefficient data of the turbine; AC impedance and magnetic pole coefficient data of the generator rotor; and oil temperature and gas concentration data in the transformer. External environmental parameters of the hydropower unit are also acquired, including: air temperature, ground temperature, relative humidity, and average wind speed data. Power grid load parameters are acquired, including: operating voltage, active power, and reactive power. Then, a data table is constructed based on the acquisition time of each parameter; the data table is shown in Table 1.

[0069]

[0070] Table 1 Data Table

[0071] In one embodiment of the present invention, determining the initial dataset further includes:

[0072] First, determine the proportion of missing values ​​in the data table relative to the total sample size. If the proportion is greater than a threshold, use the statistical imputation method, i.e., use the mean of corresponding data in the same category to fill in the missing values. If the proportion is less than the threshold, use the deletion method to directly remove the missing values. When there are many missing samples and the sample size is small, imputation can ensure a certain number of samples, ensuring the number of parameters in the dataset and the accuracy of the training model. When there are few missing samples, the sample size itself has some redundancy, and direct deletion can reduce the additional computational cost of imputation.

[0073] Next, outlier processing is performed. For samples without hydropower unit faults, outliers are deleted directly. For samples with hydropower unit faults, outliers are retained.

[0074] Then, the data is standardized using the Z-score method, and its transformation function is:

[0075]

[0076] Where μ is the mean of all sample data; σ is the standard deviation of all sample data, x * The value is the standardized value, and x is the value before processing.

[0077] Finally, feature extraction was performed using the out-of-bag data feature permutation method to obtain the initial dataset.

[0078] In one embodiment of the present invention, constructing an ARIMA model, a random forest model, and an LSTM neural network model for parameter prediction includes:

[0079] Parameters with linear characteristics are selected from the initial dataset to construct and train an ARIMA model, which is then used for parameter prediction. For example, the output power of a water turbine is directly proportional to the inlet water pressure and typically exhibits linear characteristics within its normal operating range. Parameters with nonlinear characteristics are selected from the initial dataset to construct and train a random forest model, which is then used for parameter prediction. For example, the gas concentration in an oil tank typically exhibits nonlinear characteristics, influenced by multiple factors such as temperature, pressure, and oil quality. Parameters with complex characteristics are selected from the initial dataset to construct and train an LSTM neural network model, which is then used for parameter prediction. For example, the swing coefficient is related to the structure and operating conditions of equipment such as water turbines and is difficult to describe using linear or nonlinear models, exhibiting complex characteristics. Single prediction models often focus on handling the linear or nonlinear characteristics of parameters, making it difficult to identify all relationships. Combining multiple models to construct a combined state prediction model can achieve better results than a single model.

[0080] The ARIMA model consists of an autoregressive model and a moving average model, which incorporates differencing operations. Its initial coefficients include the number of autoregressive terms, i.e. the order of the free regression model, as well as the number of moving average terms.

[0081] The random forest model averages the predictions from multiple CART regression trees before outputting the results. In this invention, the number of regression trees selected is between 100 and 200. Bootstrap sampling is selected as the sampling method to construct each basic decision tree, and the minimum number of samples is set to 1% of the data size.

[0082] The LSTM neural network model has an input dimension of 9, an output dimension of 1, and 15 hidden neurons. It is optimized using the Adam algorithm with stochastic gradient descent and the cross-entropy function as the loss function. This invention uses the TensorFlow deep learning library to build the LSTM prediction model, and the backend uses the TensorFlow-GPU open-source framework developed in C++. The data is divided according to a certain ratio, with 4000 sets of feature data as training set samples and 120 sets of data as test set samples. During the LSTM model construction, the prediction performance under different numbers of layers is compared to select an appropriate number of hidden layers and determine the relevant model parameters. The specific parameters are: 15 neurons, stride of 10, learning rate of 0.01, and Dropout ratio of 0.5.

[0083] The weights of the ARIMA model, random forest model, and LSTM neural network model are assigned using a goodness-of-fit algorithm. First, the coefficient of determination p of the goodness-of-fit algorithm is calculated. The normal range of p is [0,1]. The calculation formula is as follows:

[0084]

[0085] in, Predict the value for the i-th parameter. z is the average of the true values. i Let be the true value of the i-th parameter, and n be the total number of predicted parameters.

[0086] Then, the tangent function is used to optimize the coefficient of determination p, thereby amplifying the difference in prediction results, and finally the optimized formula for calculating the combined weights is obtained:

[0087]

[0088] Among them, w j z represents the combined weight of the j-th sub-model. j h(z) represents the p-value of the j-th sub-model. j ) represents the p value of the j-th sub-model after transformation using the tangent function.

[0089] Based on the weight value w j The weighted average of the prediction parameters of the ARIMA model, the random forest model, and the LSTM neural network model is calculated to obtain the final output result, which is used as the prediction dataset.

[0090] like Figure 2 As shown, in one embodiment of this application, it further includes:

[0091] Various types of hydropower unit faults are defined as a space Θ{F1, F2, ..., F10} as an identification framework.

[0092] The Support Vector Machine (SVM) model is constructed by using a grid search method to roughly select the penalty parameter c and the kernel function parameter g; then, the K-fold cross-validation method is used to select the optimal c and g values ​​as the optimal parameters of the SVM model.

[0093] The XGBoost model is constructed by continuously updating the number of base learners, learning rate, and maximum tree depth using an adaptive particle swarm optimization algorithm to determine the optimal parameters.

[0094] The initial warning results include: using the optimized support vector machine (SVM) model and XGBoost model, with the prediction dataset as input, to classify and warn of faults in hydropower units, and obtain two initial warning results respectively;

[0095] The final state warning result is as follows: The weighted evidence method is used to fuse the information of the two initial warning results to obtain the final state warning result of the hydropower unit.

[0096] The K-fold cross-validation method is as follows:

[0097] The K-fold cross-validation method is used to select the optimal c and g values. Specifically, the initial dataset is divided into k subsets, with k-1 subsets used as training subsets and 1 subset used as test subsets. A set of c and g values ​​is selected for training the Support Vector Machine (SVM) model, and the mean squared error (MSE) of the test results is calculated. The MSE of multiple sets of different c and g values ​​is calculated, and the optimal c and g values ​​are selected based on the principle of minimizing the MSE. In this invention, K=4, i.e., a 4-fold cross-validation method, is used. First, the data samples from the hydropower unit are classified into 4 subsets. The first 3 subsets are used as the training set, and the last subset is used for testing. This process is repeated multiple times, and an MSE is obtained each time. The data is trained 10 times, and the data obtained after each training is sorted and analyzed. The average of the 10 MSEs is then taken. Finally, the c and g values ​​are selected as the optimal parameters for SVM modeling based on the principle of minimizing the MSE.

[0098] The adaptive particle swarm optimization algorithm is as follows:

[0099] The inertia weight w for the adaptive particle swarm optimization algorithm is determined by the following formula:

[0100]

[0101] Among them, w max and w win These represent the maximum and minimum values ​​of the inertia weight, respectively, where k is the current iteration number. max This represents the maximum number of iterations.

[0102] The inertia weight ω is a key parameter for balancing the local and global searches of the algorithm. A larger ω indicates stronger global search capability but is prone to overshooting the optimal solution, while a smaller ω indicates stronger local search capability but is prone to getting trapped in local optima. Based on the evolutionary and aggregation degrees of the particle swarm, the inertia weight of the adaptive particle swarm optimization algorithm is dynamically updated to optimize the particle motion process in the search space. The improved adaptive inertia weight update formula is:

[0103] w * =w in -(w in -w min )*evol+(w max -w in )*aggr

[0104] Among them, w * w represents the adaptive inertia weight. in This represents the initial inertia weight value, where evol and aggr represent the degree of evolution and the degree of aggregation, respectively.

[0105] The formula for the degree of evolution is:

[0106]

[0107] In the formula, evol represents the evolutionary degree of the particle swarm in the range [0, 1]; e is the natural constant; c k and c k-1 Let represent the changes in fitness brought about by the k-th and (k-1)-th iterations, respectively. When the change in fitness is larger than the previous one, it indicates that there is still considerable room for improvement in the global optimum, and the evolutionary degree evol will take a smaller value. Furthermore, it is stipulated that c... k-1 When c = 0, evol reaches its minimum value of 0; conversely, evol reaches its minimum value of 1; specifically, when c = 0... k-1 =0 and c k When = 0, the position of the global optimal solution remains unchanged for multiple consecutive iterations. It is stipulated that evol is 1 at this time, and the evolution degree of the particle swarm reaches its maximum.

[0108] The formula for the degree of aggregation is:

[0109]

[0110] Here, aggr represents the degree of aggregation of the particle swarm in the range [0, 1]; N represents the number of particles in each subspace (N1, N2, N3…Nn), and the base of the log function is equal to the number of subspaces n. When all particles are evenly distributed among each subspace, i.e., N1 = N2 = N3 = …Nn, aggr reaches its minimum value of 0, at which point the degree of aggregation of the particle swarm is the lowest; when all particles are in the same subspace, i.e., there exists Ni = N, aggr reaches its maximum value of 1, at which point the degree of aggregation of the particle swarm is the highest.

[0111] The weighted evidence method is as follows:

[0112] In the two initial warning results, for each type of hydropower unit fault, there are two evidence bodies E1 and E2 to be combined, with corresponding basic probability allocation functions m1 and m2, and corresponding focal elements Ap and Bq, respectively. The common focal element of Ap and Bq is C, and the conflict coefficient between m1 and m2 is H. Then the fusion formula is:

[0113]

[0114] in, The average support level of the evidence sources for the focal element is denoted as 0. When it equals 0, it indicates that the output of one of the diagnostic models does not belong to the fault space of the hydropower unit.

[0115] By using a fusion formula, early warning results for different types of hydropower unit faults are obtained through fusion.

[0116] In one embodiment of this application, the performance of the hydropower unit's status early warning is further evaluated by calculating an average accuracy index, wherein the average accuracy index includes: average accuracy, precision, recall, and harmonic mean F1-score, wherein:

[0117] Average accuracy = (True positives + True negatives) / Total number of samples;

[0118] Precision is the proportion of samples correctly predicted as positive out of all samples predicted as positive. It measures the model's accuracy among all positive samples and is calculated using the following formula:

[0119] Precision = True positives / (True positives + False positives);

[0120] Recall rate refers to the proportion of samples correctly predicted as positive out of all truly positive samples. It measures the percentage of positive samples identified by the model out of the total number of positive samples, and is calculated using the following formula:

[0121] Recall rate = True negatives / (True negatives + False negatives);

[0122] The F1 score is the harmonic mean of precision and recall. It takes into account both the model's accuracy and recall, and is used to balance the relationship between the two. The calculation formula is as follows:

[0123] F1 score = 2 * (precision * recall) / (precision + recall).

[0124] One embodiment of the present invention discloses a joint optimization system based on peak-shaving power equipment combination prediction, comprising:

[0125] Data acquisition module: Acquires internal operating parameters, external environmental parameters, and power grid load parameters of the hydropower unit, and constructs data tables;

[0126] Data preprocessing module: Determines the initial dataset based on the data table;

[0127] Parameter prediction module: Based on the linear, nonlinear and complex features of the data in the initial dataset, construct ARIMA model, random forest model and LSTM neural network model to predict parameters; calculate the combined weights, perform a weighted average of the predicted parameters, and obtain the final output result as the prediction dataset;

[0128] Status warning module: Defines various types of hydropower unit faults into a fault space, and builds support vector machine (SVM) and XGBoost models respectively based on the initial dataset. Optimization algorithms are used to optimize the parameters of the two models. Based on the prediction dataset and fault space, the module classifies and issues warnings for hydropower unit faults, obtaining initial warning results. The initial warning results are then fused to output the final status warning result.

[0129] In one embodiment of the present invention, the parameter prediction module includes:

[0130] ARIMA model unit: Select parameters with linear features from the initial dataset, construct and train the ARIMA model, and use the trained ARIMA model to predict parameters.

[0131] Random forest model unit: Select parameters with nonlinear characteristics from the initial dataset, construct and train a random forest model, and use the trained random forest model to predict parameters;

[0132] LSTM neural network model unit: Select parameters with complex features from the initial dataset, construct and train an LSTM neural network model, and use the trained LSTM neural network model to predict parameters;

[0133] Combined weight unit: The weights of the ARIMA model, random forest model, and LSTM neural network model are assigned using a goodness-of-fit algorithm. Based on the weights, the weighted average of the prediction parameters of the ARIMA model, random forest model, and LSTM neural network model is calculated to obtain the final output result, which serves as the prediction dataset.

[0134] In one embodiment of the present invention, the status warning module includes:

[0135] SVM model unit: The grid optimization method is used to roughly select the penalty parameter c and the kernel function parameter g; then the K-fold cross-validation method is used to select the optimal c and g values ​​as the optimal parameters of the support vector machine SVM model. The prediction dataset is used as input to classify and warn of faults in hydropower units.

[0136] XGBoost model unit: The adaptive particle swarm optimization algorithm is used to continuously update the number of base learners, learning rate and maximum tree depth of the XGBoost model to determine the optimal parameters. The prediction dataset is used as input to classify and warn of faults in hydropower units.

[0137] Early warning result fusion unit: By using the weighted evidence method, the initial early warning results obtained from the SVM model unit and the XGBoost model unit are fused to obtain the final state early warning result of the hydropower unit.

[0138] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0139] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A joint optimization method based on peak-shaving power equipment combination prediction, characterized in that, The specific steps are as follows: Obtain the internal operating parameters, external environmental parameters, and power grid load parameters of the hydropower unit, and construct a data table; Based on the data table, determine the initial dataset; Based on the linear, nonlinear, and complex features of the data in the initial dataset, ARIMA, Random Forest, and LSTM neural network models are constructed for parameter prediction. Calculate the combined weights, perform a weighted average of the prediction parameters, and obtain the final output result as the prediction dataset; Multiple types of hydropower unit faults are defined as a fault space. Based on the initial dataset, a Support Vector Machine (SVM) model and an XGBoost model are built, and the parameters of the two models are optimized using an optimization algorithm. According to the prediction dataset and the fault space, the faults of the hydropower units are classified and early warnings are issued, and initial early warning results are obtained. The initial early warning results are then fused to output the final state early warning results. Specifically, the two initial early warning results are fused using a weighted evidence method to obtain the hydropower unit state early warning results. The weighted evidence method is specifically as follows: Let each hydropower unit fault in the two initial warning results have two evidence bodies E1 and E2 to be combined, with corresponding basic probability allocation functions m1 and m2, corresponding focal elements Ap and Bq, respectively, the common focal element of Ap and Bq being C, and the conflict coefficient between m1 and m2 being H. Then the fusion formula is: in, The average support level of the evidence sources for the focal element is 0, which means that the output of one of the diagnostic models does not belong to the fault space of the hydropower unit. By using a fusion formula, early warning results for different types of hydropower unit faults are obtained through fusion.

2. The joint optimization method based on peak-shaving power equipment combination prediction according to claim 1, characterized in that, The internal operating parameters include: water turbine inlet pressure, vibration, and swing coefficient data; generator rotor AC impedance and magnetic pole coefficient data; and transformer oil temperature and gas concentration data. The external environmental parameters include: air temperature, ground temperature, relative humidity, and average wind speed data. The power grid load parameters include: operating voltage, active power, and reactive power. A data table is constructed based on the time obtained from each parameter.

3. The joint optimization method based on peak-shaving power equipment combination prediction according to claim 1, characterized in that, The determination of the initial dataset includes: First, determine the proportion of missing values ​​in the data table relative to the total sample size. If the proportion is greater than a threshold, use imputation to fill in the missing values ​​in the dataset; if it is less than the threshold, use deletion to directly delete the missing values. Next, handle outliers. Then, use the Z-score method to standardize the data. Finally, use out-of-bag feature permutation to extract features and obtain the initial dataset.

4. The joint optimization method based on peak-shaving power equipment combination prediction according to claim 1, characterized in that, The construction of ARIMA model, random forest model, and LSTM neural network model for parameter prediction includes: Parameters with linear features are selected from the initial dataset, an ARIMA model is constructed and trained, and the trained ARIMA model is used for parameter prediction; parameters with nonlinear features are selected from the initial dataset, a random forest model is constructed and trained, and the trained random forest model is used for parameter prediction; parameters with complex features are selected from the initial dataset, an LSTM neural network model is constructed and trained, and the trained LSTM neural network model is used for parameter prediction. The calculation of combined weights involves weighted averaging of the prediction parameters to obtain the final output, including: The weights of the ARIMA model, random forest model, and LSTM neural network model are assigned using a goodness-of-fit algorithm. First, the coefficient of determination p of the goodness-of-fit algorithm is calculated. The normal range of p is [0,1]. The calculation formula is as follows: in, Predict the value for the i-th parameter. The average of the true values. Let be the true value of the i-th parameter, and n be the total number of predicted parameters; Then, the tangent function is used to optimize the coefficient of determination p, and finally the optimized formula for calculating the combined weights is obtained: in, Represents the combined weight of the j-th sub-model. The p-value represents the j-th sub-model. This represents the p-value of the j-th sub-model after transformation using the tangent function; Based on weight values The weighted average of the prediction parameters of the ARIMA model, the random forest model, and the LSTM neural network model is calculated to obtain the final output result, which is used as the prediction dataset.

5. The joint optimization method based on peak-shaving power equipment combination prediction according to claim 4, characterized in that, The ARIMA model consists of an autoregressive model and a moving average model, which incorporates a difference operation. Its initial coefficients include the number of autoregressive terms and the number of moving average terms. The random forest model selects bootstrap sampling to construct each basic decision tree; The LSTM neural network model is optimized using the stochastic gradient descent Adam algorithm, with the cross-entropy function as the loss function.

6. The joint optimization method based on peak-shaving power equipment combination prediction according to claim 1, characterized in that, The fault space is Θ{F1, F2, ..., F10}, which serves as the fault identification framework for hydropower units. The support vector machine (SVM) model is constructed using a grid optimization method, with a rough selection of the penalty parameter c and the kernel function parameter g. Then, the K-fold cross-validation method is used to select the optimal c and g values ​​as the optimal parameters of the support vector machine (SVM) model. The XGBoost model is constructed by continuously updating the number of base learners, learning rate, and maximum tree depth using an adaptive particle swarm optimization algorithm to determine the optimal parameters. The initial warning results include: using optimized support vector machine (SVM) and XGBoost models, with the prediction dataset as input, classifying and warning of faults in hydropower units, and obtaining initial warning results respectively; The final state warning result is: the state warning result of the hydropower unit obtained by fusing information from the two initial warning results using a weighted evidence method; The K-fold cross-validation method is as follows: The initial dataset is divided into k subsets, of which k-1 subsets are used as training subsets and 1 subset is used as test subset. A set of values ​​of c and g is selected for training the support vector machine (SVM) model, and the mean squared error (MSE) of the test results of the test subset is calculated. The MSE of multiple sets of different values ​​of c and g is calculated, and the optimal values ​​of c and g are selected based on the principle of minimizing the MSE. The adaptive particle swarm optimization algorithm is specifically as follows: The inertia weight w for the adaptive particle swarm optimization algorithm is determined by the following formula: Among them, w max and w min These represent the maximum and minimum values ​​of the inertia weight, respectively, where k is the current iteration number. max This represents the maximum number of iterations. Based on the evolution and aggregation degree of the particle swarm, the inertia weight of the adaptive particle swarm optimization algorithm is dynamically updated. The improved adaptive inertia weight update formula is as follows: in, Indicates adaptive inertia weights, This represents the initial inertia weight value, where evol and aggr represent the degree of evolution and the degree of aggregation, respectively. The weighted evidence method is specifically as follows: Let each hydropower unit fault in the two initial warning results have two evidence bodies E1 and E2 to be combined, with corresponding basic probability allocation functions m1 and m2, corresponding focal elements Ap and Bq, respectively, the common focal element of Ap and Bq being C, and the conflict coefficient between m1 and m2 being H. Then the fusion formula is: in, The average degree of support for the focal element from the evidence sources; By using a fusion formula, early warning results for different types of hydropower unit faults are obtained through fusion.

7. The joint optimization method based on peak-shaving power equipment combination prediction according to claim 1, characterized in that, It also includes evaluating the early warning performance of hydropower units by calculating average accuracy indicators, which include: average accuracy, precision, recall and harmonic mean F1-score.

8. A joint optimization system based on peak-shaving power equipment combination prediction, characterized in that, include: Data acquisition module: Acquires internal operating parameters, external environmental parameters, and power grid load parameters of the hydropower unit, and constructs data tables; Data preprocessing module: Determines the initial dataset based on the data table; Parameter prediction module: Based on the linear, nonlinear and complex features of the data in the initial dataset, construct ARIMA model, random forest model and LSTM neural network model to predict parameters; Calculate the combined weights, perform a weighted average of the prediction parameters, and obtain the final output result as the prediction dataset; State warning module: Defines various types of hydropower unit faults into a fault space, and builds Support Vector Machine (SVM) and XGBoost models respectively based on the initial dataset. Optimization algorithms are used to optimize the parameters of both models. Based on the prediction dataset and fault space, the module classifies and issues warnings for hydropower unit faults, obtaining initial warning results. The initial warning results are then fused to output the final state warning result. Specifically, a weighted evidence method is used to fuse the two initial warning results to obtain the hydropower unit state warning result. The weighted evidence method is specifically as follows: Let each hydropower unit fault in the two initial warning results have two evidence bodies E1 and E2 to be combined, with corresponding basic probability allocation functions m1 and m2, corresponding focal elements Ap and Bq, respectively, the common focal element of Ap and Bq being C, and the conflict coefficient between m1 and m2 being H. Then the fusion formula is: in, The average support level of the evidence sources for the focal element is 0, which means that the output of one of the diagnostic models does not belong to the fault space of the hydropower unit. By using a fusion formula, early warning results for different types of hydropower unit faults are obtained through fusion.

9. A joint optimization system based on peak-shaving power equipment combination prediction according to claim 8, characterized in that, The parameter prediction module includes: ARIMA model unit: Select parameters with linear features from the initial dataset, construct and train the ARIMA model, and use the trained ARIMA model for parameter prediction. Random forest model unit: Select parameters with nonlinear characteristics from the initial dataset, construct and train a random forest model, and use the trained random forest model to predict parameters; LSTM neural network model unit: Select parameters with complex features from the initial dataset, construct and train an LSTM neural network model, and use the trained LSTM neural network model to predict parameters; Combined weight unit: The weights of the ARIMA model, random forest model, and LSTM neural network model are assigned using a goodness-of-fit algorithm. Based on the weights, the weighted average of the prediction parameters of the ARIMA model, random forest model, and LSTM neural network model is calculated to obtain the final output result, which serves as the prediction dataset.

10. A joint optimization system based on peak-shaving power equipment combination prediction according to claim 8, characterized in that, The status warning module includes: SVM model unit: The grid optimization method is used to roughly select the penalty parameter c and the kernel function parameter g; then the K-fold cross-validation method is used to select the optimal c and g values ​​as the optimal parameters of the support vector machine SVM model. The prediction dataset is used as input to classify and warn of faults in hydropower units. XGBoost model unit: The adaptive particle swarm optimization algorithm is used to continuously update the number of base learners, learning rate and maximum tree depth of the XGBoost model to determine the optimal parameters. The prediction dataset is used as input to classify and warn of faults in hydropower units. Early warning result fusion unit: By using the weighted evidence method, the initial early warning results obtained from the SVM model unit and the XGBoost model unit are fused to obtain the final state early warning result of the hydropower unit.

Images