Distributed photovoltaic bus load prediction method and system based on multi-stage data optimization

Abstract

The invention provides a distributed photovoltaic bus load prediction method and system based on multi-stage data optimization. The method comprises the steps of performing multi-stage data preprocessing and modal decomposition; optimizing parameters through a grey wolf optimization algorithm GWO; and carrying out short-term feature extraction based on an improved time convolution network (ITCN). According to the method, the technical problems of single model, poor noise immunity and insufficient dynamic adaptability are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a method and system for predicting distributed photovoltaic bus loads based on multi-stage data optimization. Background Art

[0002] With the rapid development of modern power systems, the grid structure has become increasingly complex. Especially with the continuous increase in the access to renewable energy power generation, the randomness and nonlinear characteristics of distributed photovoltaic bus loads have become increasingly obvious. This change has put forward higher accuracy and timeliness requirements for load forecasting of power systems. Traditional load forecasting methods, such as statistical time series models and classic machine learning models, are difficult to accurately capture the multi-dimensional and complex characteristics of distributed photovoltaic bus loads, which often leads to insufficient prediction accuracy and difficulty in adapting to actual application needs.

[0003] In recent years, the rapid development of deep learning and intelligent optimization algorithms has provided new solutions for distributed photovoltaic bus load prediction. Deep learning models (such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs)) are superior in mining the temporal and spatial characteristics of load data, but the prediction effect of a single model still needs to be improved when facing the variability of distributed photovoltaic bus loads. In addition, intelligent optimization algorithms (such as particle swarm optimization and genetic algorithms) have unique advantages in model parameter optimization and can improve the convergence speed and prediction accuracy of the model, but they are prone to fall into local optimal solutions in complex load environments.

[0004] The existing invention patent application document "A bus load prediction method and system considering the influence of distributed photovoltaics" with publication number CN117498311A, the existing method includes: according to the current operation mode, real-time generation of bus power load data and distributed photovoltaic output data at different times, analysis of the influence of different weather factors on bus power load and distributed photovoltaics, and establishment of power load prediction model and distributed photovoltaic prediction model respectively according to weather factor analysis results and data, inputting historical data and predicted weather factor data, and obtaining bus power load prediction results and distributed photovoltaic output prediction results respectively, and calculating the prediction results of bus load off-grid according to the connection relationship between bus and distributed photovoltaics. However, the effect of removing high-frequency noise in the above-mentioned prior art needs to be improved, and it does not fully consider the characteristics of long time series, nor enhance the ability to extract key features. In addition, the convergence and prediction accuracy of the model used in the prior art need to be improved, and it is difficult to overcome the limitations of XGBoost and genetic algorithms in dealing with complex nonlinear characteristics.

[0005] The existing invention patent application document "Bus load prediction method and system considering high proportion of distributed photovoltaic access" with publication number CN112465195A, the existing method includes: step 1, performing correlation analysis on high proportion of distributed photovoltaic data and bus load data through mutual information coefficient to obtain analysis results; step 2, forming a training data set according to the analysis results; step 3, inputting the training data set into the XGBoost model and the extreme learning machine model respectively, to obtain the first output result of the XGBoost model and the second output result of the extreme learning machine model; step 4, weighting the first output result and the second output result to obtain the prediction result. However, the simple weighted summation model fusion method adopted by the prior art has poor modeling ability for randomness and nonlinear characteristics. At the same time, the existing scheme has the possibility of falling into a local optimum.

[0006] The existing invention patent application document "Adaptive prediction method, device, and computer equipment for wind-solar bus load" with publication number CN112348235A, the existing method includes: preprocessing the load data of the wind-solar bus and the corresponding meteorological data, mapping and normalizing the load data of the preprocessed wind-solar bus and the corresponding meteorological data, and performing day and night load segmentation and integration prediction based on pattern feature matching on the load data of the mapped and normalized wind-solar bus and the corresponding meteorological data to obtain the final load prediction result for the day to be predicted, and according to the final load prediction result for the day to be predicted, the load of the wind-solar bus is predicted based on the adaptive feature weight adjustment of the particle swarm algorithm. However, the aforementioned prior art relies on the day and night load prediction method of rule segmentation, and relies on fixed rules to handle load fluctuations, which makes it difficult to fully explore the overall time series coordination characteristics of load data. The prior art uses particle swarm optimization (PSO) to adjust weights, which restricts the global optimization capability, and the generalization and adaptability of the model adopted by the prior art still need to be improved.

[0007] In summary, this application achieves high-precision prediction of distributed photovoltaic bus loads by combining multimodal decomposition, global optimization and deep learning technology, solves the problems of single model, poor noise resistance, and insufficient dynamic adaptability in the existing technology, and provides strong technical support for the dispatch optimization of power systems. Summary of the invention

[0008] The technical problem to be solved by the present invention is: how to solve the technical problems of single model, poor noise resistance and insufficient dynamic adaptability in the prior art.

[0009] The present invention adopts the following technical solutions to solve the above technical problems: A distributed photovoltaic bus load prediction method based on multi-stage data optimization includes:

[0010] S1. Collect multi-stage bus load data of distributed photovoltaics, pre-process the multi-stage bus load data to obtain applicable load data, perform multi-modal decomposition on the applicable data, and obtain differential modal frequency separation data;

[0011] S2. According to the global optimal search strategy, the root mean square error RMSE is minimized to optimize the hyperparameter combination of the preset TCN model to obtain the improved temporal convolutional network ITCN. The gray wolf optimization algorithm GWO is used to optimize the parameter combination of the improved temporal convolutional network ITCN to obtain a suitable hyperparameter combination.

[0012] S3. Based on the improved temporal convolutional network ITCN, short-term feature extraction is performed according to the applicable hyperparameter combination; wherein, the improved temporal convolutional network ITCN includes: causal convolution and dilated convolution, so that the factor of the dilated convolution grows exponentially to capture the long time series characteristics, and the distributed photovoltaic bus load prediction results are obtained based on them.

[0013] Based on the existing technology, this application proposes a distributed photovoltaic bus load prediction method based on successive variational mode decomposition (SVMD), grey wolf optimization algorithm (GWO), and improved temporal convolutional network (ITCN), which overcomes the problems of insufficient model accuracy and noise resistance, difficulty in fully capturing data features, and poor real-time performance in the existing technology.

[0014] Specifically, the present invention proposes a load forecasting method based on the combination of successive variational mode decomposition (SVMD), gray wolf optimization algorithm (GWO), and improved temporal convolutional network (ITCN). Successive variational mode decomposition can decompose complex load signals into multiple frequency components, thereby reducing the non-stationarity and noise interference of multi-stage data; the gray wolf optimization algorithm can globally optimize the model parameters and enhance the convergence performance of the model; the improved temporal convolutional network is good at extracting the short-term time series characteristics of multi-stage load data, while the gated recurrent unit has an outstanding effect on modeling long-term and short-term dependencies. Based on the combined application of the above methods, the present invention can better capture the time series characteristics of multi-stage load data and realize accurate prediction of load fluctuations, thereby providing strong technical support for the dispatch optimization of power systems. It shows high accuracy and robustness in processing random fluctuations and nonlinear changes in multi-stage load data.

[0015] In a more specific technical solution, S1 includes:

[0016] S11, normalizing the multi-stage data of bus load to obtain normalized data, removing noise and abnormal values ​​from the normalized data to obtain applicable load data;

[0017] S12 performs successive variational modal decomposition (SVMD) on the applicable load data to decompose the applicable load data into natural modal components (IMFs) of no less than two different frequencies:

[0018]

[0019] In formula (2), IMF i (t) represents the i-th modal component, n is the mode number, and r(t) is the residual term.

[0020] The present invention cleans and preprocesses the original data of distributed photovoltaic bus load to ensure the integrity and accuracy of the data. The present invention uses the successive variational mode decomposition (SVMD) method to decompose the original load data into several modal components with different frequency characteristics, eliminate high-frequency noise and non-stationary factors in the load data, thereby obtaining a more stable low-frequency component, and providing high-quality data input for subsequent modeling.

[0021] In a more specific technical solution, the hyperparameter combination in S2 includes: convolution kernel size and number of convolution layers.

[0022] In a more specific technical solution, S2 includes:

[0023] S21, setting the root mean square error RMSE as the objective function to measure the deviation between the predicted value and the actual value of the preset TCN model;

[0024] S22. Use the Grey Wolf Optimization algorithm GWO to update the position of the candidate solution and perform a global search for the parameter combination of the improved temporal convolutional network ITCN to minimize the root mean square error RMSE.

[0025] In order to further improve the prediction accuracy of the model, the present invention introduces the gray wolf optimization algorithm (GWO) to optimize the important parameters in the model. The GWO algorithm simulates the hunting behavior of wolves and uses the cooperation and competition of different individuals in the group to achieve global optimization. It has the characteristics of fast convergence speed and good optimization effect. By optimizing the parameters of the ITCN model through GWO, the prediction accuracy and convergence performance of the model can be significantly improved.

[0026] In a more specific technical solution, in S21, the objective function is set using the following logic:

[0027]

[0028] In formula (3), y j is the actual load value, is the predicted value of the model, and N is the total number of samples.

[0029] The present invention compares the prediction results of the SVMD-GWO-ITCN model with the actual load data, and uses the root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R 2 The model of the present invention can significantly reduce the load prediction error, adapt to the random fluctuation and nonlinear change requirements of distributed photovoltaic bus load, and the prediction results have high accuracy in multi-stage data of loads with different fluctuation amplitudes.

[0030] In more specific technical solutions, S3 includes:

[0031] S31. According to the applicable hyperparameter combination, the network structure of the improved temporal convolutional network ITCN network is constructed;

[0032] S32, use dilated convolution to expand the receptive field and perform dilated convolution operations to capture long-term series dependencies;

[0033] S33, replacing the causal expansion convolution in the improved temporal convolutional network TCN with a two-dimensional convolution according to a preset ratio to process the time dimension and external dimension of the applicable load data;

[0034] S34. In the improved temporal convolutional network TCN, an attention mechanism is introduced to perform feature processing operations.

[0035] In a more specific technical solution, in S32, the following logic is used to express the dilated convolution:

[0036]

[0037] In formula (4), x t represents the value of the input sequence at time step t, f k is the convolution kernel weight, K is the convolution kernel size, and d is the expansion factor;

[0038] The residual module is expressed using the following logic:

[0039] o=A(x+G(x)) (5)

[0040] In formula (5), A(x) represents the activation function and G(x) represents a series of residual module transformations.

[0041] After obtaining the optimized parameters, the present invention uses an improved temporal convolutional network (ITCN) to extract time series features from load data. ITCN can effectively capture the short-term change pattern of load data through causal convolution and dilated convolution operations, avoiding information leakage while ensuring the processing capability of long-sequence multi-stage data. The features extracted by ITCN include the change trend of distributed photovoltaic bus load in the time series, providing important time-related information for the load forecasting model.

[0042] In a more specific technical solution, in S33, according to the following logic, two-dimensional convolution is used to capture sequence dependencies along the time dimension; at each time step, different feature channels are combined:

[0043]

[0044] In formula (6), x ts represents the value of the input data at time step t and other dimensions s, f ij is the weight of the two-dimensional convolution kernel, and d is the expansion factor.

[0045] In a more specific technical solution, in S34, it is assumed that the output sequence of the improved temporal convolutional network TCN is: H = [h1, h2, ..., hT];

[0046] The attention weight matrix is ​​calculated using the self-attention mechanism, and the output sequence H is weighted summed using the following logic:

[0047]

[0048] In equation (7), Q=HW Q , K=HW K 、V=HW V , W Q , W K , W V is the learnable projection matrix, d k is the scaling factor.

[0049] In a more specific technical solution, the distributed photovoltaic bus load prediction system based on multi-stage data optimization includes:

[0050] The preprocessing and multi-modal decomposition module is used to collect multi-stage data of bus load of distributed photovoltaic, preprocess the multi-stage data of bus load to obtain applicable load data, and perform multi-modal decomposition on the applicable data to obtain differential modal frequency separation data;

[0051] The model parameter tuning module is used to minimize the root mean square error RMSE according to the global optimal search strategy, optimize the hyperparameter combination of the preset TCN model, and obtain the improved temporal convolutional network ITCN. The gray wolf optimization algorithm GWO is used to optimize the parameter combination of the improved temporal convolutional network ITCN to obtain a suitable hyperparameter combination;

[0052] The photovoltaic bus load prediction module is used to extract short-term features based on the improved temporal convolutional network ITCN according to the applicable hyperparameter combination; wherein, the improved temporal convolutional network ITCN includes: causal convolution and dilated convolution, so that the factor of the dilated convolution grows exponentially to capture the long time series characteristics, and the distributed photovoltaic bus load prediction results are obtained based on this. The photovoltaic bus load prediction module is connected with the model parameter tuning module and the preprocessing and multimodal decomposition module.

[0053] Compared with the existing solutions that rely on a single XGBoost model, the present invention uses SVMD to perform modal decomposition on multi-stage load data, effectively removes high-frequency noise, and combines ITCN with dilated convolution to capture long time series characteristics, while adding an attention mechanism to enhance the ability to extract key features. In addition, by introducing GWO for global hyperparameter optimization, the convergence and prediction accuracy of the model are further improved, overcoming the limitations of XGBoost and genetic algorithms in dealing with complex nonlinear characteristics.

[0054] Compared with the method based on the linear combination of XGBoost and extreme learning machine (ELM), the present invention abandons the simple weighted sum model fusion method, directly uses ITCN to process load data end-to-end, and combines the multimodal feature components provided by SVMD to significantly improve the modeling ability of randomness and nonlinear characteristics. At the same time, GWO has a global search capability that is superior to traditional optimization algorithms, avoiding falling into local optimality, and providing the model with more efficient and stable prediction performance.

[0055] Compared with the day and night load forecasting method that relies on rule segmentation, the present invention dynamically decomposes data features through SVMD, does not need to rely on fixed rules to handle load fluctuations, and fully exploits the overall time series synergy characteristics of load data. Compared with the method of using particle swarm optimization (PSO) to adjust weights, this application uses GWO for parameter optimization, which has stronger global optimization capabilities and further improves the generalization and adaptability of the model.

[0056] Compared with the prior art, the present invention has the following advantages:

[0057] The present invention proposes a load forecasting method based on the combination of successive variational mode decomposition (SVMD), gray wolf optimization algorithm (GWO), and improved temporal convolutional network (ITCN). Successive variational mode decomposition can decompose complex load signals into multiple frequency components, thereby reducing the non-stationarity and noise interference of multi-stage data; the gray wolf optimization algorithm can globally optimize the model parameters and enhance the convergence performance of the model; the improved temporal convolutional network is good at extracting the short-term time series characteristics of multi-stage load data, while the gated recurrent unit has an outstanding effect on modeling long-term and short-term dependencies. Based on the combined application of the above methods, the present invention can better capture the time series characteristics of multi-stage load data and realize accurate prediction of load fluctuations, thereby providing strong technical support for the dispatch optimization of power systems. It shows high accuracy and robustness in processing random fluctuations and nonlinear changes in multi-stage load data.

[0058] The present invention cleans and preprocesses the original data of distributed photovoltaic bus load to ensure the integrity and accuracy of the data. The present invention uses the successive variational mode decomposition (SVMD) method to decompose the original load data into several modal components with different frequency characteristics, eliminate high-frequency noise and non-stationary factors in the load data, thereby obtaining a more stable low-frequency component, and providing high-quality data input for subsequent modeling.

[0059] In order to further improve the prediction accuracy of the model, the present invention introduces the gray wolf optimization algorithm (GWO) to optimize the important parameters in the model. The GWO algorithm simulates the hunting behavior of wolves and uses the cooperation and competition of different individuals in the group to achieve global optimization. It has the characteristics of fast convergence speed and good optimization effect. By optimizing the parameters of the ITCN model through GWO, the prediction accuracy and convergence performance of the model can be significantly improved.

[0060] The present invention compares the prediction results of the SVMD-GWO-ITCN model with the actual load data, and uses the root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R 2 The model of the present invention can significantly reduce the load prediction error, adapt to the random fluctuation and nonlinear change requirements of distributed photovoltaic bus load, and the prediction results have high accuracy in multi-stage data of loads with different fluctuation amplitudes.

[0061] After obtaining the optimized parameters, the present invention uses an improved temporal convolutional network (ITCN) to extract time series features from load data. ITCN can effectively capture the short-term change pattern of load data through causal convolution and dilated convolution operations, avoiding information leakage while ensuring the processing capability of long-sequence multi-stage data. The features extracted by ITCN include the change trend of distributed photovoltaic bus load in the time series, providing important time-related information for the load forecasting model.

[0062] The present invention solves the technical problems existing in the prior art of single model, poor noise resistance and insufficient dynamic adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a schematic diagram of the basic steps of the method for predicting distributed photovoltaic bus load based on multi-stage data optimization according to Embodiment 1 of the present invention;

[0064] Figure 2 This is a schematic diagram of data flow processing of a distributed photovoltaic bus load prediction method based on multi-stage data optimization according to Embodiment 1 of the present invention;

[0065] Figure 3 This is a schematic diagram of specific steps of data preprocessing and modal decomposition in Example 1 of the present invention;

[0066] Figure 4 This is a schematic diagram of the gray wolf position update principle of Example 1 of the present invention;

[0067] Figure 5 This is a schematic diagram of specific steps of parameter optimization in Example 1 of the present invention;

[0068] Figure 6 This is a schematic diagram of the structure of an improved temporal convolutional network ITCN according to Embodiment 1 of the present invention;

[0069] Figure 7 Schematic diagram of specific steps of short-term feature extraction in Example 1 of the present invention. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are 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.

[0071] Example 1

[0072] like Figure 1 and Figure 2As shown, the distributed photovoltaic bus load prediction method based on multi-stage data optimization provided by the present invention includes the following basic steps:

[0073] S1, multi-stage data preprocessing and modal decomposition;

[0074] like Figure 3 As shown, in this embodiment, the step S1 of data preprocessing and modal decomposition further includes the following specific steps:

[0075] S11, multi-stage data cleaning and normalization processing;

[0076] In this embodiment, the distributed photovoltaic bus load multi-stage data is cleaned and normalized to remove noise and outliers. The normalization formula is as follows, mapping the original multi-stage data X(t) to the [-1,1] interval:

[0077]

[0078] In formula (1), X represents the original load multi-stage data, X min and X max are the minimum and maximum values ​​of the multi-stage load data, respectively, and X′ is the normalized multi-stage data.

[0079] S12 Successive Variational Mode Decomposition SVMD;

[0080] In this embodiment, successive variational mode decomposition (SVMD) performs multi-modal decomposition on the load data, so that the frequency components of different modes can be separated, thereby reducing the influence of noise and extracting the periodic trend characteristics of the load.

[0081] In this embodiment, SVMD is used to decompose the normalized data C′(t) into multiple intrinsic modal components (IMFs) of different frequencies to reduce high-frequency noise and extract the main trend of the data. The decomposition process is as follows:

[0082]

[0083] In formula (2), IMF i (t) represents the i-th modal component, n is the modal number, and r(t) is the residual term. Through decomposition, the complex fluctuations of the data can be decomposed into multiple stable components; through decomposition, the high-frequency noise and low-frequency trends in the load data can be effectively separated, making the decomposed components more stable and providing better data input for subsequent modeling.

[0084] S2, optimize parameters using the Grey Wolf Optimization Algorithm GWO;

[0085] like Figure 4As shown, in this embodiment, the gray wolf optimization algorithm GWO optimizes the hyperparameter combination of the TCN model by simulating the hunting behavior of wolves, adopts a global optimal search strategy, and determines the optimal parameters with the goal of minimizing RMSE.

[0086] like Figure 5 As shown, in this embodiment, the parameter optimization step S2 also includes the following specific steps:

[0087] S21, objective function setting;

[0088] In this embodiment, the root mean square error RMSE is set as the objective function to measure the deviation between the model prediction value and the actual value;

[0089] In this embodiment, the objective function formula is as follows:

[0090]

[0091] In formula (3), y j is the actual load value, is the predicted value of the model, and N is the total number of samples. By minimizing RMSE, GWO optimizes the model's convolution kernel size, number of convolution layers and other parameters to ensure optimal prediction accuracy.

[0092] S22, GWO search and optimization;

[0093] In this embodiment, GWO is used to perform a global search for the parameter combination of ITCN to minimize the RMSE. The GWO algorithm updates the position of the candidate solution by simulating the hunting behavior of wolves, and the search for the optimal solution is based on the collaborative update of alpha, beta and delta individuals. In each iteration, by calculating the RMSE of the candidate solution, the parameter combination that minimizes the error is selected, thereby optimizing the parameters of ITCN such as the convolution kernel size and the number of convolution layers.

[0094] S3, short-term feature extraction based on improved temporal convolutional network ITCN;

[0095] like Figure 6 As shown, in this embodiment, the improved temporal convolutional network ITCN includes: causal convolution and dilated convolution. Specifically, the dilated convolution factor grows exponentially, so that the improved temporal convolutional network ITCN can effectively capture long time series features without increasing computational complexity;

[0096] like Figure 7 As shown, in this embodiment, the step S3 of short-term feature extraction also includes the following specific steps:

[0097] S31. Build ITCN network structure;

[0098] In this embodiment, the hyperparameters optimized by GWO are used to construct an improved temporal convolutional network (ITCN), which consists of multiple layers of causal convolution and dilated convolution. Causal convolution ensures the temporal causality of the model, that is, the output at each moment depends only on the previous input data to avoid information leakage.

[0099] S32, performing a dilated convolution operation;

[0100] In this embodiment, the receptive field is expanded by dilated convolution to capture long time series dependencies. The calculation formula of dilated convolution is:

[0101]

[0102] In formula (4), x t represents the value of the input sequence at time step t, f k is the convolution kernel weight, K is the convolution kernel size, and d is the dilation factor. The setting of the dilation factor d can expand the receptive field, so that ITCN can capture the long sequence dependency characteristics in the load data without increasing the complexity of the model. By setting d, ITCN can effectively extract the short-term change characteristics of the load data.

[0103] In this embodiment, the residual module can be expressed as:

[0104] o=A(x+G(x)) (5)

[0105] In formula (5), A(x) represents the activation function and G(x) represents a series of residual module transformations.

[0106] S33, the causal expansion convolution in TCN is partially replaced by a two-dimensional convolution, so that the time dimension and the external dimension of the load data can be processed simultaneously;

[0107] In this embodiment, the external dimensions include but are not limited to: temperature, humidity, and historical load. The two-dimensional convolution not only captures the sequence dependency along the time dimension, but also combines different feature channels at each time step. Its formula is:

[0108]

[0109] In formula (6), x ts represents the value of the input data at time step t and other dimensions s, f ij is the weight of the 2D convolution kernel, and d is the dilation factor. This setting allows TCN to capture the interaction pattern of load data on different features.

[0110] S34. Introducing the attention mechanism in TCN can help the model focus more on key time steps and important features, and improve prediction accuracy.

[0111] In this embodiment, the TCN output sequence is assumed to be H = [h1, h2, ..., hT], the self-attention mechanism is used to calculate the attention weight matrix, and H is weighted and summed:

[0112]

[0113] In equation (7), Q=HW Q , K=HW K 、V=HW V , W Q , W K , W V is the learnable projection matrix, d k is the scaling factor. The output after attention enhancement is H'=Attention(Q,K,V), so that the model can adaptively adjust the importance of each time step according to the correlation between time steps.

[0114] In summary, the present invention proposes a load forecasting method based on the combination of successive variational mode decomposition (SVMD), gray wolf optimization algorithm (GWO), and improved temporal convolutional network (ITCN). Successive variational mode decomposition can decompose complex load signals into multiple frequency components, thereby reducing the non-stationarity and noise interference of multi-stage data; the gray wolf optimization algorithm can globally optimize the model parameters and enhance the convergence performance of the model; the improved temporal convolutional network is good at extracting the short-term time series characteristics of multi-stage load data, while the gated recurrent unit has an outstanding effect on modeling long-term and short-term dependencies. Based on the combined application of the above methods, the present invention can better capture the time series characteristics of multi-stage load data and realize accurate prediction of load fluctuations, thereby providing strong technical support for the dispatch optimization of power systems. It shows high accuracy and robustness in processing random fluctuations and nonlinear changes in multi-stage load data.

[0115] The present invention cleans and preprocesses the original data of distributed photovoltaic bus load to ensure the integrity and accuracy of the data. The present invention uses the successive variational mode decomposition (SVMD) method to decompose the original load data into several modal components with different frequency characteristics, eliminate high-frequency noise and non-stationary factors in the load data, thereby obtaining a more stable low-frequency component, and providing high-quality data input for subsequent modeling.

[0116] In order to further improve the prediction accuracy of the model, the present invention introduces the gray wolf optimization algorithm (GWO) to optimize the important parameters in the model. The GWO algorithm simulates the hunting behavior of wolves and uses the cooperation and competition of different individuals in the group to achieve global optimization. It has the characteristics of fast convergence speed and good optimization effect. By optimizing the parameters of the ITCN model through GWO, the prediction accuracy and convergence performance of the model can be significantly improved.

[0117] The present invention compares the prediction results of the SVMD-GWO-ITCN model with the actual load data, and uses the root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R 2 The model of the present invention can significantly reduce the load prediction error, adapt to the random fluctuation and nonlinear change requirements of distributed photovoltaic bus load, and the prediction results have high accuracy in multi-stage data of loads with different fluctuation amplitudes.

[0118] After obtaining the optimized parameters, the present invention uses an improved temporal convolutional network (ITCN) to extract time series features from load data. ITCN can effectively capture the short-term change pattern of load data through causal convolution and dilated convolution operations, avoiding information leakage while ensuring the processing capability of long-sequence multi-stage data. The features extracted by ITCN include the change trend of distributed photovoltaic bus load in the time series, providing important time-related information for the load forecasting model.

[0119] The present invention solves the technical problems existing in the prior art of single model, poor noise resistance and insufficient dynamic adaptability.

[0120] 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 aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A distributed photovoltaic bus load prediction method based on multi-stage data optimization, characterized in that: The method comprises: S1. Collecting multi-stage data of bus load of distributed photovoltaic, pre-processing the multi-stage data of bus load to obtain applicable load data, performing multi-modal decomposition on the applicable data to obtain differential modal frequency separation data; S2. According to the global optimal search strategy, the root mean square error RMSE is minimized, the hyperparameter combination of the preset TCN model is optimized, and an improved temporal convolutional network ITCN is obtained. The parameter combination of the improved temporal convolutional network ITCN is optimized by using the gray wolf optimization algorithm GWO to obtain a suitable hyperparameter combination; S3. Based on the improved temporal convolutional network ITCN, short-term feature extraction is performed according to the applicable hyperparameter combination; wherein the improved temporal convolutional network ITCN includes: causal convolution and dilated convolution, so that the factor of the dilated convolution grows exponentially to capture long time series characteristics, and the distributed photovoltaic bus load prediction results are obtained based on them.

2. The distributed photovoltaic bus load prediction method based on multi-stage data optimization according to claim 1 is characterized in that: The S1 includes: S11, normalizing the multi-stage data of the bus load to obtain normalized data, removing noise and abnormal values ​​from the normalized data to obtain the applicable load data; S12 performs successive variational modal decomposition (SVMD) on the applicable load data to decompose the applicable load data into natural modal components (IMFs) of no less than two different frequencies: In formula (2), IMF i (t) represents the i-th modal component, n is the mode number, and r(t) is the residual term.

3. The distributed photovoltaic bus load prediction method based on multi-stage data optimization according to claim 1 is characterized in that: The hyperparameter combination in S2 includes: convolution kernel size and number of convolution layers.

4. The distributed photovoltaic bus load prediction method based on multi-stage data optimization according to claim 1 is characterized in that: The S2 includes: S21, setting the root mean square error RMSE as the objective function to measure the deviation between the predicted value and the actual value of the preset TCN model; S22. Use the grey wolf optimization algorithm GWO to update the position of the candidate solution, and perform a global search on the parameter combination of the improved temporal convolutional network ITCN to minimize the root mean square error RMSE.

5. The distributed photovoltaic bus load prediction method based on multi-stage data optimization according to claim 4 is characterized in that: In S21, the objective function is set using the following logic: In formula (3), y j is the actual load value, is the predicted value of the model, and N is the total number of samples.

6. The distributed photovoltaic bus load prediction method based on multi-stage data optimization according to claim 1 is characterized in that: The S3 includes: S31, constructing the network structure of the improved temporal convolutional network ITCN network according to the applicable hyperparameter combination; S32, using the dilated convolution to expand the receptive field, performing a dilated convolution operation, and capturing long-term series dependencies; S33, replacing the causal dilation convolution in the improved temporal convolutional network TCN with a two-dimensional convolution according to a preset ratio to process the time dimension and external dimension of the applicable load data; S34. In the improved temporal convolutional network TCN, an attention mechanism is introduced to perform feature processing operations.

7. The distributed photovoltaic bus load prediction method based on multi-stage data optimization according to claim 6 is characterized in that: In S32, the dilated convolution is expressed using the following logic: In formula (4), x t represents the value of the input sequence at time step t, f k is the convolution kernel weight, K is the convolution kernel size, and d is the expansion factor; The residual module is expressed using the following logic: o=A(x+G(x)) (5) In formula (5), A(x) represents the activation function and G(x) represents a series of residual module transformations.

8. The distributed photovoltaic bus load prediction method based on multi-stage data optimization according to claim 6 is characterized in that: In S33, according to the following logic, the two-dimensional convolution is used to capture sequence dependency along the time dimension; at each time step, different feature channels are combined: In formula (6), x ts represents the value of the input data at time step t and other dimensions s, f ij is the weight of the two-dimensional convolution kernel, and d is the expansion factor.

9. The distributed photovoltaic bus load prediction method based on multi-stage data optimization according to claim 6 is characterized in that: In S34, the output sequence of the improved temporal convolutional network TCN is assumed to be: H = [h1, h2, ..., hT]; The attention weight matrix is ​​calculated using the self-attention mechanism, and the output sequence H is weighted summed using the following logic: In formula (7), Q=HW Q , K=HW K 、V=HW V , W Q , W K , W V is the learnable projection matrix, d k is the scaling factor.

10. The distributed photovoltaic bus load prediction system is optimized based on multi-stage data, characterized in that: The system comprises: A preprocessing and multi-modal decomposition module is used to collect multi-stage data of bus load of distributed photovoltaic, preprocess the multi-stage data of bus load to obtain applicable load data, and perform multi-modal decomposition on the applicable data to obtain differential modal frequency separation data; A model parameter tuning module is used to minimize the root mean square error RMSE according to the global optimal search strategy, optimize the hyperparameter combination of the preset TCN model, obtain an improved temporal convolutional network ITCN, and use the gray wolf optimization algorithm GWO to optimize the parameter combination of the improved temporal convolutional network ITCN to obtain a suitable hyperparameter combination; A photovoltaic bus load prediction module is used to extract short-term features based on the improved temporal convolutional network ITCN and according to the applicable hyperparameter combination; wherein the improved temporal convolutional network ITCN includes: causal convolution and dilated convolution, so that the factor of the dilated convolution grows exponentially to capture long time series features, and the distributed photovoltaic bus load prediction result is obtained based on the processing, and the photovoltaic bus load prediction module is connected to the model parameter tuning module and the preprocessing and multimodal decomposition module.

Images