Electric vehicle charging load prediction method, system and storage medium combining sequence decomposition and feature extraction

Abstract

The present invention relates to the field of charging load prediction technology, and is a method, system and storage medium for predicting the charging load of electric vehicles that combines sequence decomposition and feature extraction. The method comprises the following steps: performing variational mode decomposition on time series load data to obtain multiple intrinsic mode functions; after zero-filling and expanding each intrinsic mode function, converting it to the frequency domain; enhancing the frequency domain data through a multi-layer perceptron based on complex domain calculation; constructing a prediction model based on a complex domain attention module and a residual module, inputting the enhanced spectrum data into the prediction model, extracting the data characteristic frequency combination through the complex domain attention module and the residual module, and taking the weighted sum of all intrinsic mode functions as the output of the prediction model; calculating the difference between the real data and the predicted data based on the mean square error of the time series and the mean absolute value error of the spectrum, obtaining the time domain and frequency domain mixing loss, and updating the prediction model parameters. The present invention solves the problem of large prediction errors under the influence of random load fluctuations.

Description

Technical Field

[0001] The present invention relates to the technical field of charging load prediction, and in particular to a method, system and storage medium for predicting the charging load of an electric vehicle combining sequence decomposition and feature extraction. Background Art

[0002] Power system operation requires a balance between power supply and demand to maintain system stability and reliability. However, the widespread adoption of electric vehicles (EVs) has become a major trend in electricity consumption, and the random nature of their charging needs has introduced uncertainty on the demand side of the power system, significantly increasing the difficulty of power dispatch. Therefore, power systems require accurate load forecasting to facilitate advance generation planning and ensure a balanced supply and demand.

[0003] Power load forecasting involves predicting power load for the next several hours or days based on historical data. Traditional time series forecasting methods rely on statistical regression models. However, regression models ignore the complex interactions between multiple independent variables and are not well suited for modeling long-term dependencies. Deep learning models are better able to handle long-term and complex time series relationships. Common neural network architectures suitable for time series forecasting include multi-layer perceptrons (MLPs), convolutional networks, recurrent neural networks, and transformer-based networks. These models all capture the correlation between different time steps from a time series perspective, using the autocorrelation of the sequence to establish a connection between historical and future data for prediction.

[0004] A core issue in long-term time series forecasting is that methods should be able to expressively capture long-term dependencies. However, the aforementioned methods are all limited to forecasting the time series itself, which can easily lead to local overfitting. For a time series, frequency contains periodic information. While high frequencies can be used to extract short-term dependencies, low frequencies can also be used to extract long-term dependencies. Frequency-based neural networks extract the spectral characteristics of the load series, first predicting the spectrum of the future series, and then converting it back to the time domain through an inverse Fourier transform.

[0005] Due to the random nature of charging behavior, load data is complex and noisy in both the time and frequency domains. This data characteristic increases the difficulty for neural network models to extract effective features and make accurate predictions. To improve the generalization performance of deep learning models, it is necessary to select appropriate feature variables to enrich the input of the prediction model. Summary of the Invention

[0006] In response to the defects of the existing technology that load data is random and noisy, the present invention proposes an electric vehicle charging load prediction method, system and storage medium that combine sequence decomposition and feature extraction. Variational mode decomposition (VMD) is used to generate multiple subsequences with clear features, and then the frequency feature extraction is used to predict the spectrum. This solves the technical problems of large errors in prediction results caused by the neglect of long-term dependencies by time series models and the prediction deviation caused by noise.

[0007] In an embodiment of the present invention, a method for predicting electric vehicle charging load by combining sequence decomposition and feature extraction includes the following steps:

[0008] S1. Obtain the operating data of the charging station, standardize the operating data, and divide it into a training set, a validation set, and a test set according to the proportion. Perform variational mode decomposition on the operating data of each set to generate multiple intrinsic mode functions;

[0009] S2, padding each intrinsic mode function with zeros to extend the data length, and then using fast Fourier transform to obtain spectrum data with higher frequency resolution;

[0010] S3. Use data slicing to divide each spectrum data into multiple continuous frequency bands and then splice them together to convert the one-dimensional data into two-dimensional data. Input the two-dimensional data into a multi-layer perceptron based on complex domain calculation to enhance the spectrum data.

[0011] S4. Constructing a prediction model based on the complex domain attention module and the residual module, inputting the enhanced spectrum data into the constructed prediction model, extracting the data characteristic frequency combination through the complex domain attention module and the residual module, and taking the weighted sum of all intrinsic mode functions as the output of the prediction model;

[0012] S5. Predict a period of historical data so that the future series will be consistent with the historical data in terms of data trend; calculate the difference between the real data and the predicted data based on the mean square error of the time series and the mean absolute error of the spectrum to obtain the time-domain and frequency-domain mixed loss; use this mixed loss as the loss function to update the prediction model parameters and realize the training of the prediction model.

[0013] In an embodiment of the present invention, a system for predicting electric vehicle charging load that combines sequence decomposition and feature extraction, based on the above-mentioned prediction method, includes the following modules:

[0014] The intrinsic mode function generation module obtains the operating data of the charging station, standardizes the operating data, and divides it into training, validation, and test sets according to the proportion. The operating data of each set is subjected to variational mode decomposition to generate multiple intrinsic mode functions.

[0015] The spectrum data generation module pads each intrinsic mode function with zeros to expand the data length, and then uses fast Fourier transform to obtain spectrum data with higher frequency resolution;

[0016] The spectrum enhancement module uses data slicing to divide each spectrum data into multiple continuous frequency bands and then splices them together, converting the one-dimensional data into two-dimensional data. The two-dimensional data is input into a multi-layer perceptron based on complex domain calculation to enhance the spectrum data.

[0017] The prediction model construction module builds a prediction model based on the complex domain attention module and the residual module. The enhanced spectrum data is input into the constructed prediction model, and the characteristic frequency combination of the data is extracted through the complex domain attention module and the residual module. The weighted sum of all intrinsic mode functions is used as the output of the prediction model;

[0018] The prediction model training module predicts a period of historical data so that the future series will be consistent with the historical data in terms of data trend. The module calculates the difference between the real data and the predicted data based on the mean square error of the time series and the mean absolute error of the spectrum to obtain the time-domain and frequency-domain mixed loss. The mixed loss is used as the loss function to update the prediction model parameters to realize the training of the prediction model.

[0019] In an embodiment of the present invention, there is also a storage medium on which computer instructions are stored. When the computer instructions are executed by a processor, the steps of the prediction method in the embodiment of the present invention are implemented.

[0020] It can be seen from the above technical solutions that the present invention has the following advantages:

[0021] 1. The present invention predicts future time series from the frequency domain, constructs an MLP and attention structure based on complex operations, uses the attention structure to extract characteristic frequency combinations from frequency domain data, and uses MLP to enhance mid- and high-frequency data in the block-wise frequency domain data because the attention mechanism focuses more on low frequencies.

[0022] 2. This invention performs preprocessing by decomposing and transforming the data based on its time-frequency characteristics. It uses variational mode decomposition (VMD) to generate IMFs with clearer frequency domain characteristics and eliminate noise. Based on the characteristics of the discrete Fourier transform, it extends the sequence length to generate higher-resolution frequency domain data. VMD can decompose complex time series into multiple modal components with clear frequency components, effectively removing noise and simplifying the data structure, thereby improving the prediction accuracy of deep learning models.

[0023] 3. The loss function of the present invention combines the MSE in the time domain and the MAE in the frequency domain to more accurately capture the loss of trends and cycles. During training, a period of historical data is predicted and the loss of historical data is used to enhance the prediction of data distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of a method for predicting electric vehicle charging load by combining sequence decomposition and feature extraction in an embodiment of the present invention;

[0025] Figure 2 The prediction model framework in the embodiment of the present invention is mainly composed of three sub-components: 1) VMD, which decomposes the sequence and reduces noise; 2) the frequency attention module, which extracts the frequency features of the sequence; 3) the time-frequency mixed loss, which updates the prediction model from both the time domain and the frequency domain;

[0026] Figure 3 Schematic diagram of frequency enhancement structure using complex domain MLP;

[0027] Figure 4 The complex field computation graph for linear mapping;

[0028] Figure 5 The following is a visualization of the comparison results of input 96 and prediction 96, where (a) is the experimental result of the prediction method of this embodiment, (b) is the experimental result of the ATFNet algorithm, (c) is the experimental result of the FEDformer algorithm, (d) is the experimental result of the PatchTST algorithm, (e) is the experimental result of the Autoformer algorithm, and (f) is the experimental result of the Nonstationary Transformer algorithm.

[0029] Figure 6 The following is a visualization of the input 96-prediction 96 ablation results, where (a) is the load prediction result of the frequency attention module only, (b) is the load prediction result of FAM plus TF loss, and (c) is the load prediction result of this embodiment that combines VMD, FAM and TF loss. DETAILED DESCRIPTION

[0030] The technical solutions of the present invention are further described in detail below with reference to the embodiments and accompanying drawings. It is apparent that the embodiments described below are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0031] Example

[0032] See also Figure 1 、 Figure 2 This embodiment provides a charging load prediction method combining signal decomposition and frequency feature extraction, comprising the following steps:

[0033] S1. Obtain the operating data of the charging station, standardize the operating data, and divide it into a training set, a validation set, and a test set according to the proportion. Perform variational mode decomposition on the operating data of each set to generate multiple intrinsic mode functions.

[0034] In this embodiment, the operating data is charging power data; performing variational modal decomposition on the operating data specifically includes the following steps:

[0035] S11. Divide the charging power data into samples according to a preset division standard to generate a training set, a validation set, and a test set. Standardize the charging power data in each set to obtain standardized sample data.

[0036] The dataset contains charging load data from an electric vehicle charging station at an airport in a certain city. The station has 55 charging piles with a rated power of 120 kW. The charging load data used in the experiment was collected from April 1 to May 31, 2023. To show detailed changes in the charging load, the sampling interval was set to 15 minutes.

[0037] The dataset is divided into training set, validation set and test set according to the ratio of 0.5, 0.25 and 0.25. The data of each set is standardized respectively. The standardization formula is:

[0038]

[0039] In the formula, h(t) is the original data set data, t represents the data time, and the maximum length is N. μ is the data mean. σ is the data standard deviation. ′ (t) is the standardized sample data.

[0040] S12. Perform variational modal decomposition on the standardized sample data to generate multiple modal functions corresponding to the sample data.

[0041] That is, the variational mode decomposition algorithm is used to extract the signal features from the standardized sample data. Set the sample data h ′ (t) is decomposed into K modal functions (i.e. modal components), each modal function u k (t) are all kept at the center frequency ω k around.

[0042] S13. An initial constraint model is constructed with the optimization objectives of minimizing the estimated bandwidth of all modal functions and making the sum of all modal functions equal to the corresponding sample data.

[0043] This step minimizes the sum of the estimated bandwidths of all modal functions while satisfying the constraint that the sum of all modal functions is equal to the original input signal (i.e., sample data). The initial constraint model is constructed as follows:

[0044]

[0045] Where * represents the convolution operation, u k (t) represents the mode function; δ(t) in the brackets is the Dirac function, which does not change the mode function when convolved with it, and 1 / πt is convolved with the mode function to perform the Hilbert transform; i represents the imaginary unit, and the Dirac function and 1 / πt are used as the real and imaginary parts, respectively, and summed to form the analytical signal; Indicates the partial derivative of t, K is the number of preset modal functions, {u1,u2,…,u k} is the standardized sample data h ′ (t) each mode function, {ω1,ω2,…,ω k} is the center frequency of each mode function.

[0046] S14. Use a preset Lagrangian expansion expression to update the initial constraint model to generate a target constraint model; through the target constraint model, use an alternating direction multiplier algorithm and a preset convergence formula to iteratively update the modal function to generate multiple intrinsic mode functions.

[0047] S2. Each intrinsic mode function is padded with zeros to extend the data length, and then a fast Fourier transform is used to obtain spectrum data with higher frequency resolution.

[0048] In this embodiment, the following steps are specifically included:

[0049] S21, according to the historical data length L and the predicted data length H, fill zero at the end of each intrinsic mode function until the data length is Get the corresponding extended mode function.

[0050] For the intrinsic mode function u(t) of length L, H zeros are added after the sequence data to expand the intrinsic mode function u(t) to a length of The extended modal function u ext (t).

[0051] S22. Convert each extended mode function from the time domain to the frequency domain using fast Fourier transform.

[0052] Among them, the formula of fast Fourier transform is:

[0053]

[0054] Where F[k] is the spectrum, index k is the frequency point, and there are a total of discrete frequencies. ext (t) is the extended mode function, index n is the time, there are a total of The last H terms are all zero, so they can be ignored when calculating the spectrum, which speeds up the calculation. i θ =cosθ+isinθ is a complex exponential function. As k changes, A series of functions with a period of L + H are generated. Each discrete value of the input sequence is multiplied by the corresponding complex exponential function, and all these products are summed to obtain a frequency weight in the output spectrum.

[0055] In addition, since the front and back segments of the spectrum have a conjugate property, this embodiment only retains half of the spectrum to reduce the time complexity in subsequent calculations:

[0056]

[0057] In the formula, the kth frequency and the penultimate kth frequency in the spectrum are in a conjugate relationship, and only half of the data is retained, thereby reducing both time and space complexity.

[0058] S3. Use data slicing to divide each spectrum data into multiple continuous frequency bands and then splice them together to convert the one-dimensional data into two-dimensional data; input the two-dimensional data into a multi-layer perceptron based on complex domain calculation to enhance the spectrum data.

[0059] Specifically, if Figure 3 As shown, the following steps are included:

[0060] S31. Split the spectrum of the intrinsic mode function into n continuous sub-spectrum graphs by using a non-overlapping data block method, and stack the n segments of data to form two-dimensional data.

[0061] S32. For each continuous sub-spectrum graph, perform normalization processing according to its own mean and variance.

[0062] The formula for standardization is:

[0063]

[0064] Where f[n,i] is two-dimensional data, subscript n represents the number of sub-spectrum segments, and there are N segments in total; subscript i represents the sequence number of the frequency in each spectrum segment, and its length is p; μ n is the average value of each spectrum; σ n is the standard deviation of each spectrum segment.

[0065] S33. Splice the two-dimensional data back into one-dimensional data according to the original spectrum order; then input the standardized data into a multi-layer perceptron based on complex domain calculation for frequency enhancement.

[0066] Figure 4 It is the complex field computation graph of linear mapping, and the specific formula is:

[0067] X out =X·W=(X re W re -X im W im )+i(X im W re +X re W im )

[0068] Where, X is the input spectrum data, X re and X im are the real and imaginary parts of the spectrum data respectively; W is the neural network weight parameter, W re and W im are the real and imaginary parts of the weight respectively; X out The result of complex field multiplication.

[0069] Among them, the multi-layer perceptron structure based on complex domain calculation specifically includes a normalization layer, a fully connected layer I, a gelu activation layer, a dropout layer I, a fully connected layer II and a dropout layer II connected in sequence.

[0070] S4. Construct a prediction model based on the complex domain attention module and the residual module, input the enhanced spectrum data into the constructed prediction model, extract the data characteristic frequency combination through the complex domain attention module and the residual module, and take the weighted sum of all eigenmode functions as the output of the prediction model.

[0071] Specifically, the following steps are included:

[0072] S41. Use linear transformation to embed the enhanced spectrum data into the high-dimensional spectrum In the example, we use three learnable parameters and Linearly map the high-dimensional spectrum into the query matrix Q h , key matrix K h Sum matrix V h .

[0073] S42. Use multi-head attention to extract different frequency combinations, perform complex dot product attention on each head, and finally concatenate the attention outputs of all heads. Use the residual network to fuse the multi-head data to obtain the fused attention output.

[0074] The attention calculation formula is:

[0075]

[0076] In the formula, head hDenotes the attention output, the subscript h denotes the number of heads, there are N attention heads, that is, there are N parallel attention modules. Each attention head uses the query matrix Q h Multiply by the key matrix K h The conjugate transpose of , and then normalize the value through the softmax function to get the attention score, and finally multiply it by the value matrix V h Get the output of each attention head.

[0077] The residual network calculation formula is:

[0078]

[0079] Where F is the concatenation of attention heads (head1,…,head H ), MLP is a multi-layer perceptron based on complex domain calculation, is the attention output after residual network fusion.

[0080] S5. Predict a period of historical data so that the future series will be consistent with the historical data in terms of data trend; calculate the difference between the real data and the predicted data based on the mean square error of the time series and the mean absolute error of the spectrum to obtain the time-domain and frequency-domain mixed loss; use this mixed loss as the loss function to update the prediction model parameters and realize the training of the prediction model.

[0081] In this embodiment, the values at time t and time t+1 are predicted at the same time at time t; the predicted value at time t is historical data, which is a case of using known data to predict known data. Its prediction error will be much smaller, so this part of the predicted data is added when training the model to reduce the training error.

[0082] By mixing the time domain and frequency domain losses and using them as the loss function to update the prediction model parameters, the cycle and trend characteristics of the charging load can be accurately captured.

[0083] Specifically, the following steps are included:

[0084] S51. Output the fused attention through inverse Fourier transform Convert to time domain signal

[0085] S52, the length of the tag data Y is H, but the time domain signal The length is When training the prediction model, the label data is supplemented with a historical sequence of length L to make the label data consistent with the length of the time domain signal to help predict data trends.

[0086] S53. Calculate the difference between the real data and the predicted data based on the mean square error of the time series and the mean absolute error of the spectrum to obtain a time-domain and frequency-domain mixed loss; use the mixed loss as a loss function to update the prediction model parameters to achieve training of the prediction model.

[0087] In this embodiment, the calculation formula of the mixing loss is:

[0088]

[0089] Where α is the loss weight; f i is the frequency spectrum of the labeled data, is the spectrum of the predicted data, the subscript i represents the i-th frequency, and there are n frequencies in total; y i is the time series of label data, To predict the time series of data, the subscript i represents the i-th moment, and there are n moments in total.

[0090] S6. Use the trained prediction model to predict the charging load of electric vehicles.

[0091] To verify the superiority of the method proposed in this example, five methods, including Nonstationary Transformer, PatchTST, and Autoformer for time domain prediction, and ATFNet and FEDformer for frequency domain prediction, were used to train and test the same training and test samples. The prediction effect was evaluated using mean absolute error and mean square error:

[0092]

[0093] Among them, MAE is the mean absolute error, MSE is the mean square error, n is the number of moments to be predicted, and y i and are the actual and predicted values of the charging load at the charging station at time i. The experimental results are shown in Table 1.

[0094] Table 1 Comparative experimental results

[0095]

[0096] For both short-term and long-term predictions, this embodiment achieves the lowest error across both metrics, significantly improving upon other models. When the prediction length is 24, the MSE and MAE decrease by 25.6% and 14.2%, respectively, compared to the control model with the lowest error. Furthermore, across various prediction lengths, the model in this embodiment is the most stable, with the error decreasing slowly as the prediction length increases.

[0097] Figure 5A sample of the test set is shown, (a) is the experimental result of the prediction method of this embodiment, (b) is the experimental result of the ATFNet algorithm, (c) is the experimental result of the FEDformer algorithm, (d) is the experimental result of the PatchTST algorithm, (e) is the experimental result of the Autoformer algorithm, and (f) is the experimental result of the Nonstationary Transformer algorithm.

[0098] from Figure 5 It can be seen that the data set has a certain periodicity, in which the amplitude fluctuates between cycles, whether it is the high-frequency or low-frequency part, which shows that the noise is distributed in the entire spectrum. The performance of the model predicted in the time domain is generally not as good as the model in the frequency domain. The period identified by the time domain model is single, and the global dependency of the data is almost not captured. Moreover, the local features are not well identified, and the entire predicted data fluctuates smoothly. In contrast, the frequency domain models can capture the periodicity very well, and the three frequency domain models can fit the curve of the real data, but there are certain differences in the processing of local dependencies. The model proposed in this embodiment uses VMD to process data, and while decomposing the modal components (i.e., modal functions) with clearer frequency characteristics, it also eliminates the noise in the entire frequency band, so the model of this embodiment is also closer to the true value in amplitude.

[0099] Table 2 Ablation experiment results

[0100]

[0101] The present invention mainly consists of three modules: 1. Frequency Attention Module (FAM), 2. Time-Frequency Mixed Loss (TF loss), and 3. VMD parallel structure. In order to verify the effectiveness of the proposed method, this embodiment designed three models for ablation comparison. Model 1 only has FAM, model 2 is FAM plus TF loss, and model 3 is a combination of FAM, TF loss and VMD, that is, the complete model proposed by the present invention. The three models are trained and tested on the same training samples and test samples. The evaluation indicators of the prediction effect are the mean absolute error and the mean square error. As can be seen from Table 2, all three models increase the accuracy of model prediction.

[0102] Figure 6 The distribution of data in the dataset can illustrate the role of different modules. Figure 6 Figure (a) shows the load forecasting result with only the frequency attention module. The frequency attention module can capture the periodicity of the data, and the overall curve fluctuations are consistent with the true value. However, the overall value of the prediction in Figure (a) is higher than the true data. Figure 6 Figure (b) shows the FAM plus TF loss, which is used to supplement the prediction of data trends. The result is that the trend of the predicted values is more consistent with the real data. Figure 6 As shown in Figure (c), VMD, IMF, and TF loss are combined. VMD is first used to obtain k IMFs, which are then applied in parallel to k FAMs. Finally, the results are summed and the TF loss is used to update the model. The decomposed IMFs have clearer frequency domain information. The decomposition process also eliminates noise across the entire frequency band. The parallel VMD structure allows for targeted prediction of characteristic frequencies, resulting in optimal prediction results for both high and low frequencies.

[0103] Based on the same inventive concept and on the basis of the aforementioned prediction method, this embodiment further provides an electric vehicle charging load prediction system that combines sequence decomposition and feature extraction, including the following modules:

[0104] The intrinsic mode function generation module obtains the operating data of the charging station, standardizes the operating data, and divides it into training, validation, and test sets according to the proportion. The operating data of each set is subjected to variational mode decomposition to generate multiple intrinsic mode functions.

[0105] The spectrum data generation module pads each intrinsic mode function with zeros to expand the data length, and then uses fast Fourier transform to obtain spectrum data with higher frequency resolution;

[0106] The spectrum enhancement module uses data slicing to divide each spectrum data into multiple continuous frequency bands and then splices them together, converting the one-dimensional data into two-dimensional data. The two-dimensional data is input into a multi-layer perceptron based on complex domain calculation to enhance the spectrum data.

[0107] The prediction model construction module builds a prediction model based on the complex domain attention module and the residual module. The enhanced spectrum data is input into the constructed prediction model, and the characteristic frequency combination of the data is extracted through the complex domain attention module and the residual module. The weighted sum of all intrinsic mode functions is used as the output of the prediction model;

[0108] The prediction model training module predicts a period of historical data so that the future series will be consistent with the historical data in terms of data trend. The module calculates the difference between the real data and the predicted data based on the mean square error of the time series and the mean absolute error of the spectrum to obtain the time-domain and frequency-domain mixed loss. The mixed loss is used as the loss function to update the prediction model parameters to realize the training of the prediction model.

[0109] The above modules are respectively used to implement steps S1-S5 of the above prediction method. When the training of the prediction model is completed, the charging load of the electric vehicle can be predicted by the prediction model.

[0110] In addition, this embodiment further provides a storage medium on which computer instructions are stored. When the computer instructions are executed by a processor, the steps of the prediction method described in this embodiment are implemented.

[0111] The present invention solves the problem of large prediction errors under the influence of random load fluctuations, and can achieve lower errors even when data fluctuates violently. Moreover, the prediction error changes stably under different prediction lengths, and the prediction results have higher prediction accuracy and stability.

[0112] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A method for predicting electric vehicle charging load by combining sequence decomposition and feature extraction, characterized in that: The following steps are involved: S1. Obtain operating data from the charging station, standardize the data, and divide it into a training set, a validation set, and a test set in proportion. Perform variational mode decomposition on the operating data of each set to generate multiple intrinsic mode functions. S2, padding each intrinsic mode function with zeros to extend the data length, and then using fast Fourier transform to obtain spectrum data with higher frequency resolution; S3. Use data slicing to divide each spectrum data into multiple continuous frequency bands and then splice them together to convert the one-dimensional data into two-dimensional data. Input the two-dimensional data into a multi-layer perceptron based on complex domain calculation to enhance the spectrum data. S4. Constructing a prediction model based on the complex domain attention module and the residual module, inputting the enhanced spectrum data into the constructed prediction model, extracting the data characteristic frequency combination through the complex domain attention module and the residual module, and taking the weighted sum of all intrinsic mode functions as the output of the prediction model; S5. Predict a period of historical data to ensure that the future series is consistent with the historical data in terms of data trend. Calculate the difference between the real data and the predicted data based on the mean square error of the time series and the mean absolute error of the spectrum to obtain the time-domain and frequency-domain mixing loss. The mixed loss is used as the loss function to update the prediction model parameters to achieve the training of the prediction model; Step S4 includes: S41, use linear transformation to embed the enhanced spectrum data into the high-dimensional spectrum, and then use three learnable parameters and Linearly map the high-dimensional spectrum into the query matrix Q h , key matrix K h Sum matrix V h ; S42. Use multi-head attention to extract different frequency combinations, perform complex dot product attention on each head, and finally concatenate the attention outputs of all heads. Use the residual network to fuse the multi-head data to obtain the fused attention output.

2. The prediction method according to claim 1, characterized in that The operating data in step S1 is charging power data; performing variational modal decomposition on the operating data includes the following steps: S11. Divide the charging power data into samples according to a preset division standard to generate a training set, a validation set, and a test set. Standardize the charging power data in each set to obtain standardized sample data. S12. performing variational modal decomposition on the standardized sample data to generate multiple modal functions corresponding to the sample data; S13, constructing an initial constraint model with the optimization goal of minimizing the estimated bandwidth of all modal functions and making the sum of all modal functions equal to the corresponding sample data; S14. Use a preset Lagrangian expansion expression to update the initial constraint model to generate a target constraint model; through the target constraint model, use an alternating direction multiplier algorithm and a preset convergence formula to iteratively update the modal function to generate multiple intrinsic mode functions.

3. The prediction method according to claim 2, characterized in that The initial constraint model is constructed as: Where t represents the data moment, * represents the convolution operation, and u k (t) represents the mode function; δ(t) in the brackets is the Dirac function, which does not change the mode function when convolved with the mode function, and 1 / πt performs the Hilbert transform when convolved with the mode function; i represents the imaginary unit, and the Dirac function and 1 / πt are summed as the real and imaginary parts respectively to form the analytical signal; Indicates the partial derivative of t, K is the number of preset modal functions, {u1,u2,…,u k } are the modal functions of the standardized sample data h′(t), {ω1,ω2,…,ω k } is the center frequency of each mode function.

4. The prediction method according to claim 1, wherein: Step S2 includes: S21, according to the historical data length L and the predicted data length H, fill zero at the end of each intrinsic mode function until the data length is Get the corresponding extended mode function; S22. Convert each extended mode function from the time domain to the frequency domain using fast Fourier transform.

5. The prediction method according to claim 1, wherein: Step S3 includes: S31, splitting the spectrum of the intrinsic mode function into n continuous sub-spectrum graphs by a non-overlapping data block method, and stacking the n segments of data to form two-dimensional data; S32, performing normalization processing on each continuous sub-spectrum according to its own mean and variance; S33. Splice the two-dimensional data back into one-dimensional data according to the original spectrum order; then input the standardized data into a multi-layer perceptron based on complex domain calculation for frequency enhancement.

6. The prediction method according to claim 5, characterized in that The standardization processing formula in S32 is: Where f[n,i] is two-dimensional data, subscript n represents the number of sub-spectrum segments, and there are N segments in total; subscript i represents the sequence number of the frequency in each spectrum segment, and its length is p; μ n is the average value of each spectrum; σ n is the standard deviation of each spectrum segment.

7. The prediction method according to claim 1, wherein: Step S5 includes: S51, convert the fused attention output into a time domain signal through inverse Fourier transform; S52. When training the prediction model, the label data of length H is supplemented with a historical sequence of length L, so that the length of the label data is consistent with that of the time domain signal; S53. Calculate the difference between the real data and the predicted data based on the mean square error of the time series and the mean absolute error of the spectrum to obtain the time-domain and frequency-domain mixed loss; use the time-domain and frequency-domain mixed loss as the loss function to update the prediction model parameters to achieve training of the prediction model.

8. An electric vehicle charging load forecasting system combining sequence decomposition and feature extraction, characterized in that: The prediction system is based on the prediction method according to any one of claims 1 to 7, comprising: The intrinsic mode function generation module obtains the operating data of the charging station, standardizes the operating data, and divides it into training, validation, and test sets according to the proportion. The operating data of each set is subjected to variational mode decomposition to generate multiple intrinsic mode functions. The spectrum data generation module pads each intrinsic mode function with zeros to expand the data length, and then uses fast Fourier transform to obtain spectrum data with higher frequency resolution; The spectrum enhancement module uses data slicing to divide each spectrum data into multiple continuous frequency bands and then splices them together, converting the one-dimensional data into two-dimensional data. The two-dimensional data is input into a multi-layer perceptron based on complex domain calculation to enhance the spectrum data. The prediction model construction module builds a prediction model based on the complex domain attention module and the residual module. The enhanced spectrum data is input into the constructed prediction model, and the characteristic frequency combination of the data is extracted through the complex domain attention module and the residual module. The weighted sum of all intrinsic mode functions is used as the output of the prediction model; The prediction model training module predicts a period of historical data so that the future series will be consistent with the historical data in terms of data trend. The module calculates the difference between the real data and the predicted data based on the mean square error of the time series and the mean absolute error of the spectrum to obtain the time-domain and frequency-domain mixed loss. The mixed loss is used as the loss function to update the prediction model parameters to realize the training of the prediction model.

9. A storage medium having computer instructions stored thereon, which, when executed by a processor, implements the steps of the prediction method according to any one of claims 1 to 7.

Images