Short-term power load prediction method based on SWT and improved Autoformer model

Abstract

The invention relates to a short-term power load prediction method based on an SWT and an improved Autoformer model, and the method employs the SWT to carry out the de-noising processing of original load data, effectively reduces the interference of noise on a prediction result, and provides clearer input for the model. An autocorrelation mechanism and a multi-head autocorrelation mechanism are introduced into an Autoformer model, so that the capturing capability of similar sub-processes and long-term dependency relationships in a time sequence is enhanced, and the prediction precision is remarkably improved; meanwhile, a multi-scale window moving average decomposition technology is adopted in a hybrid expert decomposition module MOED, the season component and the trend component of a load sequence are effectively extracted, and the adaptability, prediction accuracy and robustness of the model to complex data are improved; the KAN module replaces a traditional multilayer perceptron MLP, parameter consumption is reduced, the generalization ability of the model is improved, and the prediction performance of the model is further enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of power system load forecasting, and in particular to a short-term power load forecasting method based on SWT and an improved Autoformer model. Background Art

[0002] With the increasing demand for new energy technologies, the power system continues to develop in the direction of intelligence, and power dispatch planning is gradually market-oriented. Therefore, accurate short-term power load forecasting plays a vital role in the safe, economical and stable operation of the power system. Traditional methods include time series analysis, Kalman filtering, etc.

[0003] Time series analysis can use the autocorrelation in the data to predict future power loads, and can be applied to load forecasts of various time scales, from short-term to medium- and long-term. At the same time, many software and programming libraries provide a large number of time series analysis tools, making this model easy to implement. However, time series analysis requires that the data is stationary, with a constant mean, variance, and autocovariance, but actual power load data is often non-stationary and needs to be converted into stationary data through methods such as differentiation and smoothing, which will cause certain data loss. At the same time, the time series model is sensitive to outliers in the data, which has a greater impact on the prediction results. In addition, the prediction interval of the time series analysis method is too wide, which reduces the practicality of the prediction.

[0004] The Kalman filter is a recursive algorithm that does not require the storage of a large amount of historical data. It only requires the estimated value and covariance of the current data to make predictions. Due to its recursive nature, it can quickly update the prediction results and is suitable for real-time or online predictions. However, the traditional Kalman filter is based on a linear system, while the power load system is usually highly nonlinear, which limits the application scope of the Kalman filter. In addition, the Kalman filter is very sensitive to the selection of the initial state estimate and the covariance matrix. Inappropriate initial values ​​will cause the filter to converge slowly or diverge.

[0005] Traditional methods for predicting power load can effectively process linear data, but the prediction results for nonlinear data are not accurate. In recent years, load forecasting methods have begun to adopt machine learning and deep learning technologies. The Transformer model has developed rapidly in natural language processing and computer vision, and has obvious advantages in predicting power load by capturing long-distance dependencies in sequence data through the self-attention mechanism. However, the standard Transformer model faces high computational load, internal limitations and low efficiency in processing long sequences in long sequence prediction. Therefore, the Autoformer model emerged, which replaces the self-attention mechanism in the Transformer model with the autocorrelation mechanism, and replaces the point-by-point attention with the segment-by-segment attention. This change enhances the ability of the Autoformer to capture sequence fragments more effectively and reduces the time complexity of improving long-term sequence prediction. At the same time, the Autoformer model uses a decomposition technology based on a fixed window sliding average as a submodule in its multi-layer prediction model. This decomposition technology effectively solves the time cost challenge faced by traditional decomposition algorithms in real scenarios. However, due to the inherent distribution bias in unstable load sequences, the decomposition accuracy of the fixed window sliding average decomposition method is low, and the Autoformer model cannot fully extract the seasonal component and trend component of the load sequence, and still cannot effectively improve the load forecasting accuracy. Summary of the invention

[0006] In view of this, the present invention provides a short-term power load forecasting method based on SWT and improved Autoformer model, which effectively solves the problem of low prediction accuracy in short-term load forecasting due to nonlinear data and abnormal situations caused by the non-stationarity of power system load.

[0007] To achieve the above object, the present invention provides a short-term power load forecasting method based on SWT and improved Autoformer model, comprising the following steps:

[0008] S1. Obtain original load data and perform preprocessing;

[0009] S2, using synchronous compression wavelet transform SWT to denoise the pre-processed load data;

[0010] S201, decomposing the original sequence of load data by using wavelet transform;

[0011] S202, setting the wavelet transform coefficient of the noise to zero;

[0012] S203, reconstructing the original load sequence using inverse wavelet transform;

[0013] S3, input the reconstructed original load sequence into the MOEAutoformer encoder;

[0014] The MOEAutoformer encoder includes a multi-head autocorrelation mechanism module, two MOED modules and a Komogorov-Arnold network KAN;

[0015] S4, using a hybrid expert decomposition MOED module based on a multi-scale moving average filter to extract seasonal terms and trend terms from the load sequence, and the seasonal terms are input into the MOEAutoformer decoder;

[0016] The MOEAutoformer decoder includes a multi-head autocorrelation mechanism module, an autocorrelation mechanism module, three MOED modules and a KAN network;

[0017] The MOED module extracts various trend items from the load sequence and combines different weights to obtain a comprehensive trend item X trend , the expression is as follows:

[0018] X trend =Softmax(L(x))*(F(x))

[0019] Among them, x represents the input load sequence, L() represents the linear function, F() represents a set of average pooling filters, and Softmax() represents the weight matrix;

[0020] Then deduct the comprehensive trend term from the original load sequence to obtain the seasonal term X seasonal , the expression is as follows:

[0021] X seasonal =xX trend

[0022] The obtained trend term and seasonal term are input into the MOEAutoformer decoder;

[0023] S5, inputting the sequence output by the MOEAutoformer encoder into the MOEAutoformer decoder as an input of the second multi-head autocorrelation mechanism module of the seasonal term;

[0024] S6. The predicted season item and the predicted trend item obtained by the MOEAutoformer decoder are added to obtain a prediction result.

[0025] Preferably, the preprocessing is to clean the original load data and convert the load sequence with outliers and noise into clearer load data. The synchronous compression wavelet transform SWT formula is as follows:

[0026] SWT f (ω,b)=∫WT f(a,b)a -3 / 2 δ[(ω,b)-ω]da

[0027] Among them, a represents the scale, b represents the displacement factor, ω represents the frequency, and δ represents the Dirac function.

[0028] Preferably, the reconstructed original load sequence is input into the MOEAutoformer encoder, the sequence obtained by the multi-head autocorrelation mechanism module is added to the input reconstructed original load sequence and input into the first MOED module, the first MOED module is connected to the KAN network, the sequence output by the KAN network is added to the sequence output by the first MOED module and input into the second MOED module to obtain the output of the MOEAutoformer encoder.

[0029] Preferably, the seasonal term is input into the MOEAutoformer decoder, the sequence output by the multi-head autocorrelation mechanism module is added to the original seasonal term and input into the first MOED module, the first MOED module is connected to the autocorrelation mechanism module, the sequence output by the autocorrelation mechanism module is added to the seasonal term output by the first MOED module and input into the second MOED module, the second MOED module is connected to the KAN network, the sequence output by the KAN network is added to the seasonal term output by the second MOED module and input into the third MOED module, and the third MOED module is decomposed to obtain the predicted seasonal term.

[0030] Preferably, the trend item is input into the MOEAutoformer decoder, and the trend item is accumulated with the trend items decomposed by each MOED module to obtain a predicted trend item.

[0031] Preferably, the multi-head autocorrelation mechanism is used to capture different patterns and relationships in the data and map them to different subspaces through different linear deformations, including the following steps:

[0032] Enter d model The dimensional matrices Q, K, V are divided into h d model / h Dimension Q i ,K i ,V i , where Q represents the query vector, K represents the value vector, and V represents the key vector;

[0033] Map Q, K, and V to different subspaces through different linear transformations;

[0034] Calculate Q i and K i The autocorrelation coefficient of

[0035] Filter out the time delay τ corresponding to the top-k autocorrelation coefficients j,j∈{1,2,3…,k};

[0036] According to the time delay τ j ,j∈{1,2,3…,k} for V i Perform time-delay aggregation;

[0037] Concatenate the h header results and perform linear transformation to generate the final output;

[0038] MultiHead(Q,K,V)=W output ×Concat(head1,head2,…,head h )

[0039] head i =Correlation(Q i ,K i ,V i ),i∈(1,2,…,h)

[0040] Among them, MultiHead() represents the output result of the multi-head autocorrelation mechanism, W output Represents trainable weights, Concat() represents the concatenation function, head i represents the i-th head, and Correlation() represents the attention mechanism calculation.

[0041] Preferably, the autocorrelation coefficient formula is as follows:

[0042]

[0043] Among them, R xx (τ) represents the discrete time series {χ t}Autocorrelation coefficient in the time domain, S xx (f) represents the discrete time series {χ t}Autocorrelation coefficient in the frequency domain, χ t and χ t-τ Represents the sequence { t}, R xx (f) represents the time domain autocorrelation coefficient after inverse fast Fourier transform, F() represents fast Fourier transform, F -1 () represents the inverse fast Fourier transform, F * () represents the inverse fast Fourier transform conjugate, t represents time, τ represents the time constant, and f represents the frequency.

[0044] Preferably, the KAN network calculation formula is as follows:

[0045]

[0046] Among them, Φ i ,i∈(0,1,…,L-1) represents the spline nonlinear function of the i-th layer, x represents the multidimensional input data, and ° represents the function combination operation.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. The present invention uses SWT to denoise the original load data, effectively reducing the interference of noise on the prediction results, improving data quality, and thus providing a clearer input for the model;

[0049] 2. By introducing the autocorrelation mechanism and the multi-head autocorrelation mechanism, the improved Autoformer model enhances the ability to capture similar sub-processes and long-term dependencies in the time series, and significantly improves the prediction accuracy;

[0050] 3. The MOED module of the present invention effectively extracts the seasonal component and trend component of the load sequence through the multi-scale window sliding average decomposition technology, improving the adaptability, prediction accuracy and robustness of the model to complex data; the KAN module replaces the traditional multi-layer perceptron MLP, reduces parameter consumption and improves the generalization ability of the model, further enhancing the prediction performance of the model;

[0051] 4. The SWT-MEOA model proposed in the present invention provides a new solution for short-term load forecasting of power systems, which is of great significance for improving the efficiency of power resource scheduling and allocation and the safe operation of power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is the framework diagram of the SWT-MOEAutoformer model of the present invention;

[0053] Figure 2 This is the original load sequence curve diagram of the present invention;

[0054] Figure 3 A comparison diagram of load sequence curves before and after noise reduction of the present invention;

[0055] Figure 4 This is a time-frequency comparison diagram of SWT before and after noise reduction of the present invention;

[0056] Figure 5 This is a diagram of the time-delay aggregation process of the autocorrelation mechanism module of the present invention;

[0057] Figure 6 This is a schematic diagram of the multi-head autocorrelation mechanism module of the present invention;

[0058] Figure 7 This is a comparison chart of KAN and MLP processing in the present invention;

[0059] Figure 8 A comparison chart of the SWT-MOEA model of the present invention and various deep learning models with the actual load value;

[0060] Fig. 9 This is a comparison chart of the absolute errors of the SWT-MOEA model of the present invention and various deep learning models;

[0061] Fig.10 A comparison chart of load forecasting results of the SWT-MOEA model of the present invention and a variety of Transformer-based models and improved models;

[0062] Fig.11 This is a relative error comparison chart of the SWT-MOEA model of the present invention and various Transformer-based models and improved models. DETAILED DESCRIPTION

[0063] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0064] In order to reduce the problem of low power load forecasting accuracy caused by nonlinear data and data anomalies, improve the efficiency of power resource scheduling and allocation, and maintain the safe operation of the power system, the present invention provides the following technical solutions: a short-term power load forecasting method based on SWT and improved Autoformer model, comprising the following steps:

[0065] S1. Obtain the original load data and preprocess it, clean the original load data, and convert the load sequence with outliers and noise into clearer load data;

[0066] S2, using synchronous compression wavelet transform SWT to denoise the pre-processed load data;

[0067] S201, decomposing the original sequence of load data by using wavelet transform;

[0068] S202, setting the wavelet transform coefficient of the noise to zero;

[0069] S203, reconstructing the original load sequence using inverse wavelet transform;

[0070] The original load data is as follows Figure 2 As shown, SWT is used for denoising to achieve high-precision and high-resolution time domain analysis. The specific calculation method is as follows:

[0071]

[0072] Where f(t) represents the signal, A k (t) represents the instantaneous amplitude of the k harmonic component, θ k (t) represents the instantaneous phase of the k harmonic component, and e(t) represents the noise;

[0073] Perform continuous wavelet transform on the signal f(t) to obtain the wavelet coefficients WT(a,b):

[0074]

[0075] Among them, a represents the scale, b represents the displacement factor, ψ() represents the wavelet basis function, is the conjugate of ψ;

[0076] The frequency domain expression of wavelet coefficient WT(a,b) is:

[0077]

[0078] in, represents the Fourier transform, represents the conjugate of the wavelet basis function;

[0079] The synchronous compression wavelet transform SWT formula is as follows:

[0080] SWT f (ω,b)=∫WT f (a,b)a -3 / 2 δ[(ω,b)-ω]da

[0081] Among them, ω represents frequency, δ represents Dirac function;

[0082] Figure 3 , Figure 4 They are curve comparison chart and time-frequency comparison chart before and after noise reduction respectively;

[0083] S3, input the reconstructed original load sequence into the MOEAutoformer encoder;

[0084] The MOEAutoformer encoder includes a multi-head autocorrelation mechanism module, two MOED modules, and a Komogorov-Arnold network KAN;

[0085] The reconstructed original load sequence is input into the MOEAutoformer encoder. The sequence obtained by the multi-head autocorrelation mechanism module is added to the input reconstructed original load sequence and input into the first MOED module. The first MOED module is connected to the KAN network. The sequence output by the KAN network is added to the sequence output by the first MOED module and input into the second MOED module to obtain the output of the MOEAutoformer encoder.

[0086] S4, using a hybrid expert decomposition MOED module based on a multi-scale moving average filter to extract seasonal terms and trend terms from the load sequence, and the seasonal terms are input into the MOEAutoformer decoder;

[0087] The MOEAutoformer decoder includes a multi-head autocorrelation mechanism module, an autocorrelation mechanism module, three MOED modules and a KAN network;

[0088] The seasonal term is input into the MOEAutoformer decoder. The sequence output by the multi-head autocorrelation mechanism module is added to the original seasonal term and input into the first MOED module. The first MOED module is connected to the autocorrelation mechanism module. The sequence output by the autocorrelation mechanism module is added to the seasonal term output by the first MOED module and input into the second MOED module. The second MOED module is connected to the KAN network. The sequence output by the KAN network is added to the seasonal term output by the second MOED module and input into the third MOED module. The third MOED module decomposes to obtain the predicted seasonal term.

[0089] The trend item is input into the MOEAutoformer decoder, and the trend item is accumulated with the trend items decomposed by each MOED module to obtain the predicted trend item;

[0090] The MOED module extracts various trend items from the load sequence and combines different weights to obtain a comprehensive trend item X trend , the expression is as follows:

[0091] X trend =Softmax(L(x))*(F(x))

[0092] Among them, x represents the input load sequence, L() represents the linear function, F() represents a set of average pooling filters, and Softmax() represents the weight matrix;

[0093] Then deduct the comprehensive trend term from the original load sequence to obtain the seasonal term X seasonal , the expression is as follows:

[0094] X seasonal =xX trend

[0095] The obtained trend term and seasonal term are input into the MOEAutoformer decoder;

[0096] S5, inputting the sequence output by the MOEAutoformer encoder into the MOEAutoformer decoder as an input of the second multi-head autocorrelation mechanism module of the seasonal term;

[0097] S6. The predicted season item and the predicted trend item obtained by the MOEAutoformer decoder are added to obtain a prediction result.

[0098] The autocorrelation mechanism in the Autoformer model includes periodic dependence and time-delay aggregation. Periodic dependence estimates the sequence period and finds similar subprocesses by the autocorrelation coefficients between subsequences. Similar subprocesses can be determined by the autocorrelation coefficients between subsequences. The larger the autocorrelation coefficient, the more similar the subsequences.

[0099] like Figure 5 As shown, the subsequence information is aggregated with time delay, and the autocorrelation coefficient is probabilized by the Softmax function. j , j∈{1,2,3…,k}, perform rolling alignment on similar subsequences, and then multiply the aligned subsequences with the corresponding probability correlation coefficients for aggregation, and use the top-k time delays τ1,τ2,…τ with the largest autocorrelation coefficients k Implements sequence-level connections.

[0100] The multi-head autocorrelation mechanism enhances the expressiveness of the model by capturing different patterns and relationships in the data, allowing each head to focus on different aspects, enhancing robustness to noise and anomalies, and accelerating training by learning different feature representations simultaneously; the flexibility of this mechanism enables it to adapt to complex real-world scenarios;

[0101] The multi-head autocorrelation mechanism module is mapped to different subspaces through different linear deformations, including the following steps:

[0102] Enter d model The dimensional matrices Q, K, V are divided into h d model / h Dimension Q i ,K i ,V i , where Q represents the query vector, K represents the value vector, and V represents the key vector;

[0103] Map Q, K, and V to different subspaces through different linear transformations;

[0104] Calculate Q i and K i The autocorrelation coefficient is as follows:

[0105]

[0106] Among them, R xx (τ) represents the discrete time series {χ t}Autocorrelation coefficient in the time domain, S xx (f) represents the discrete time series {χ t}Autocorrelation coefficient in the frequency domain, χ t and χ t-τ Represents the sequence { t}, R xx (f) represents the time domain autocorrelation coefficient after inverse fast Fourier transform, F() represents fast Fourier transform, F -1 () represents the inverse fast Fourier transform, F * () represents the inverse fast Fourier transform conjugate, t represents time, τ represents the time constant, and f represents the frequency;

[0107] Filter out the time delay τj corresponding to the top-k autocorrelation coefficient , j∈{1,2,3…,k};

[0108] According to the time delay τ j ,j∈{1,2,3…,k} for V i Perform time-delay aggregation;

[0109] Concatenate the h header results and perform linear transformation to generate the final output;

[0110] MultHead(Q,K,V)=W output ×Concat(head1,head2,…,head h )

[0111] head i =Correlation(Q i ,K i ,V i ),i∈(1,2,…,h)

[0112] Among them, MultiHead() represents the output result of the multi-head autocorrelation mechanism, W output Represents trainable weights, Concat() represents the concatenation function, head i represents the i-th head, and Correlation() represents the attention mechanism calculation;

[0113] The load time series ideally reflects the continuous real distribution, but often presents distribution changes due to discrete sensor sampling, which results in different data representations capturing different features at different sampling frequencies even if they come from the same source; distribution changes are a real phenomenon in time series, and reasonable use of this phenomenon can effectively improve the prediction accuracy of the model; the research focus of the present invention is to predict future loads of different prediction lengths based on data collected at fixed intervals; since different forecast lengths mean processing data of different time spans, these data may show different distribution characteristics, the present invention uses a multi-scale moving average filter to analyze the load series to extract trends and periodic changes, thereby improving the accuracy and robustness of long-term load forecasting;

[0114] In other Autoformer models, the fully connected layer in the multi-layer perceptron MLP is usually used to map the extracted features to the final prediction value. However, the use of the fully connected layer in the MLP will result in a large amount of non-embedded parameter consumption, which is usually difficult to explain without post-analysis. Therefore, the Kolmogorov-Arnold network KAN is used instead of MLP in the present invention; KAN is similar to MLP in structure, but the core difference is that KAN can put the learnable activation function on the weight instead of fixing it on the neuron, which allows each weight parameter to be parameterized by a one-dimensional spline function, thereby maintaining the feature learning ability of MLP while reducing parameter consumption and enhancing the generalization energy of the model; KAN fully utilizes the high-precision advantage of the spline function in low-dimensional approximation through the combination of spline functions and weights, while avoiding the dimensionality disaster that may occur in MLP in high-dimensional feature learning; in the load forecasting task, KAN can more effectively capture the nonlinear characteristics in the load data and improve the model's adaptability to complex change patterns. The KAN network calculation formula is as follows:

[0115]

[0116] Among them, Φ i ,i∈(0,1,…,L-1) represents the spline nonlinear function of the i-th layer, x represents multidimensional input data, and ° represents function combination operation;

[0117] MLP is an alternation of a linear transformation and a nonlinear transformation, and its expression is:

[0118]

[0119] Where x represents multidimensional input data, W i ,i∈(0,1,…,L-1) represents the affine transformation of the i-th layer including weights and offsets, ° represents the function combination operation, and σ represents the activation function;

[0120] like Figure 7 As shown, suppose there is a network with a depth of L, each layer has a width of N, and each spline has an order of k and is distributed in G intervals. Therefore, KAN has O(N 2 L(G+k))~O(N 2 LG) parameters. Under the same conditions, MLP has X parameters. Although MLP seems to have fewer parameters, in fact, KAN has N fewer parameters than MLP, which allows KAN to save more parameters and have better generalization effect.

[0121] In this embodiment, the original load data set used is from a district in Wuxi City. The data records the power load from January 1, 2022 to January 15, 2022. The collection frequency is once every 15 minutes, and a total of 12,400 sets of data are collected; 70% of the data are used as training sets and 30% of the data are used as test sets. Due to factors such as temperature changes, weather conditions, and user behavior, the power load data shows significant non-steady-state fluctuations; the method proposed in the present invention is compared with a variety of deep learning models, a variety of Transformer models, and improved models.

[0122] The SWT-MOEA model proposed in this paper is compared with the LSTM, GRU, TCN, CNN-LSTM, and ICEEMDAN-SSA-LSTM deep learning models. Figure 8 As shown in the figure, the prediction curve of the SWT-MOEA model proposed in the present invention almost coincides with the actual load value, showing a very high prediction accuracy; in contrast, although the LSTM and GRU models can capture the overall trend, the prediction error is large during the period of large load fluctuations; although the TCN model has a good prediction effect in smooth areas, it performs poorly when the load changes sharply; although the CNN-LSTM and ICEEMDAN-SSA-LSTM combined models have improved, they still fail to achieve the prediction accuracy of the SWT-MOEA model;

[0123] Fig. 9 The distribution of prediction errors of the proposed method and LSTM, GRU, TCN, CNN-LSTM, and ICEEMDAN-SSA-LSTM deep learning models is shown in the form of box plots; the error box plot of the SWT-MOEA model shows the smallest error range and the lowest outlier, indicating that its prediction results are more stable and reliable; while the error distribution of other deep learning models is relatively dispersed, especially when the prediction deviation is large, the error fluctuation is more obvious. These comparative results fully demonstrate the superiority of the SWT-MOEA model in deep learning models, especially in processing complex nonlinear time series data.

[0124] The SWT-MOEA model proposed in this invention is compared with Transformer, Autoformer, Informer and FEDformer. Fig.10 As shown in Figure 2, the SWT-MOEA model proposed in this invention has the highest fit with the true value. Fig.11 The mean error curve shows that the prediction error curve of the SWT-MOEA model has the smallest fluctuation, which directly proves the superiority and effectiveness of the proposed model.

[0125] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A short-term power load forecasting method based on SWT and improved Autoformer model, characterized in that: The following steps are involved: S1. Obtain original load data and perform preprocessing; S2, using synchronous compression wavelet transform SWT to denoise the pre-processed load data; S201, decomposing the original sequence of load data by using wavelet transform; S202, setting the wavelet transform coefficient of the noise to zero; S203, reconstructing the original load sequence using inverse wavelet transform; S3, input the reconstructed original load sequence into the MOEAutoformer encoder; The MOEAutoformer encoder includes a multi-head autocorrelation mechanism module, two MOED modules and a Komogorov-Arnold network KAN; S4, using a hybrid expert decomposition MOED module based on a multi-scale moving average filter to extract seasonal terms and trend terms from the load sequence, and the seasonal terms are input into the MOEAutoformer decoder; The MOEAutoformer decoder includes a multi-head autocorrelation mechanism module, an autocorrelation mechanism module, three MOED modules and a KAN network; The MOED module extracts various trend items from the load sequence and combines different weights to obtain a comprehensive trend item X trend , the expression is as follows: X trend =Soft max(L(x))*(F(x)) Among them, x represents the input load sequence, L() represents the linear function, F() represents a set of average pooling filters, and Softmax() represents the weight matrix; Then deduct the comprehensive trend term from the original load sequence to obtain the seasonal term X seasonal , the expression is as follows: X seasonal =x-X trend The obtained trend term and seasonal term are input into the MOEAutoformer decoder; S5, inputting the sequence output by the MOEAutoformer encoder into the MOEAutoformer decoder as an input of the second multi-head autocorrelation mechanism module of the seasonal term; S6. The predicted season item and the predicted trend item obtained by the MOEAutoformer decoder are added to obtain a prediction result.

2. According to claim 1, a short-term power load forecasting method based on SWT and improved Autoformer model is characterized in that: Preprocessing is to clean the original load data and convert the load sequence with outliers and noise into clearer load data. The frequency domain formula of synchronous compression wavelet transform SWT is as follows: SWT f (ω,b)=∫WT f (a,b)a -3 / 2 δ[(ω,b)-ω]da Among them, a represents the scale, b represents the displacement factor, ω represents the frequency, and δ represents the Dirac function.

3. According to claim 1, a short-term power load forecasting method based on SWT and improved Autoformer model is characterized in that: The reconstructed original load sequence is input into the MOEAutoformer encoder. The sequence obtained by the multi-head autocorrelation mechanism module is added to the input reconstructed original load sequence and input into the first MOED module. The first MOED module is connected to the KAN network. The sequence output by the KAN network is added to the sequence output by the first MOED module and input into the second MOED module to obtain the output of the MOEAutoformer encoder.

4. According to claim 1, a short-term power load forecasting method based on SWT and improved Autoformer model is characterized in that: The seasonal term is input into the MOEAutoformer decoder, and the sequence output by the multi-head autocorrelation mechanism module is added to the original seasonal term and input into the first MOED module. The first MOED module is connected to the autocorrelation mechanism module, and the sequence output by the autocorrelation mechanism module is added to the seasonal term output by the first MOED module and input into the second MOED module. The second MOED module is connected to the KAN network, and the sequence output by the KAN network is added to the seasonal term output by the second MOED module and input into the third MOED module. The third MOED module decomposes to obtain the predicted seasonal term.

5. The short-term power load forecasting method based on SWT and improved Autoformer model according to claim 1 is characterized in that: The trend item is input into the MOEAutoformer decoder, and the trend item is accumulated with the trend items decomposed by each MOED module to obtain a predicted trend item.

6. A short-term power load forecasting method based on SWT and improved Autoformer model according to claim 1, characterized in that: The multi-head autocorrelation mechanism is used to capture different patterns and relationships in the data and map them to different subspaces through different linear deformations, including the following steps: Enter d model The dimensional matrices Q, K, V are divided into h d model / h Dimension Q i ,K i ,V i , where Q represents the query vector, K represents the value vector, and V represents the key vector; Map Q, K, and V to different subspaces through different linear transformations; Calculate Q i and K i The autocorrelation coefficient of Filter out the time delay τ corresponding to the top-k autocorrelation coefficients j ,j∈{1,2,3…,k}; According to the time delay τ j ,j∈{1,2,3…,k} for V i Perform time-delay aggregation; Concatenate the h header results and perform linear transformation to generate the final output; MultiHead(Q,K,V)=W output ×Concat(head1,head2,…,head h ) head i =Correlation(Q i ,K i ,V i ),i∈(1,2,…,h) Among them, MultiHead() represents the output result of the multi-head autocorrelation mechanism, W output Represents trainable weights, Concat() represents the concatenation function, head i represents the i-th head, and Correlation() represents the attention mechanism calculation.

7. A short-term power load forecasting method based on SWT and improved Autoformer model according to claim 6, characterized in that: The autocorrelation coefficient formula is as follows: Among them, R xx (τ) represents the discrete time series {χ t }Autocorrelation coefficient in the time domain, S xx (f) represents the discrete time series {χ t }Autocorrelation coefficient in the frequency domain, χ t and χ t-τ Represents the sequence { t }, R xx (f) represents the time domain autocorrelation coefficient after inverse fast Fourier transform, F() represents fast Fourier transform, F -1 () represents inverse fast Fourier transform, F * () represents the inverse fast Fourier transform conjugate, t represents time, τ represents the time constant, and f represents the frequency.

8. The short-term power load forecasting method based on SWT and improved Autoformer model according to claim 1 is characterized in that: The KAN network calculation formula is as follows: Among them, Φ i , i∈(0,1,…,L-1) represents the spline nonlinear function of the i-th layer, x represents the multidimensional input data, Represents function combination operation.

Images