Short-term power load prediction method and system based on improved DSSFA-SAC-ConvLSTM

Abstract

The invention discloses a short-term power load prediction method and system based on improved DSSFA-SAC-ConvLSTM, and the method comprises the steps: firstly collecting and preprocessing regional historical power load data, then employing DSSFA to analyze and extract features, constructing a feature matrix, generating a distance adjacency matrix and a dynamic adjacency matrix based on a geographic distance and DSSFA features, fusing the distance adjacency matrix and the dynamic adjacency matrix into an autocorrelation matrix, and carrying out the prediction of a short-term power load based on the autocorrelation matrix. Then splicing the feature matrix and the autocorrelation matrix, inputting the spliced feature matrix and the autocorrelation matrix into a ConvLSTM model for training, adjusting a learning rate by adopting Warmup and improved OneCycleLR, and finally mapping an output feature into a power load prediction value; according to the method, key features are extracted through DSSFA, the data dimension is effectively reduced, important information is reserved, a dynamic adjacency matrix is generated in combination with a geographic distance and a self-attention mechanism, spatial information is fully considered, the accuracy of multi-region prediction is improved, the ConvLSTM learning rate is optimized by adopting a Warmup and an improved OneCycleLR strategy, the model training effect is enhanced, and the prediction efficiency is improved. And the accuracy of power load prediction is further improved, so that the power load prediction method has more excellent performance in a complex scene.

Description

Technical Field

[0001] The present invention relates to power load forecasting means, belonging to the field of power load forecasting, and particularly relates to a short-term power load forecasting method and system based on improved DSSFA-SAC-ConvLSTM. Background Technique

[0002] With the rapid development of social economy and the continuous adjustment of energy structure, the power demand shows a continuous growth trend. Especially during peak hours and holidays, the volatility of power load is significantly enhanced. The non-linear and non-stationary characteristics of this load, combined with the gradual expansion of the power grid scale and the increasing complexity of load characteristics, significantly increase the difficulty of short-term power load forecasting and pose new challenges to the applicability and accuracy of traditional forecasting methods.

[0003] In the existing short-term power load forecasting technologies, some studies adopt the method of two-dimensional variational mode decomposition combined with convolutional long short-term memory neural network. This method improves the forecasting accuracy of short-term power load to a certain extent. However, two-dimensional variational mode decomposition has deficiencies in dealing with trend components and complex noises, which limits the further improvement of the model forecasting accuracy. On the other hand, a regional short-term power load forecasting method based on the combination of support vector machine, STL time series decomposition method and long short-term memory neural network has also been proposed. Although this method shows certain advantages in time series decomposition and forecasting, it fails to fully consider the spatial information in power load data, which limits its performance in multi-region power load forecasting tasks and results in low forecasting accuracy. Summary of the Invention

[0004] The purpose of the present invention is to overcome the above-mentioned defects and problems existing in the prior art, and provide a short-term power load forecasting method and system based on improved DSSFA-SAC-ConvLSTM with relatively high accuracy.

[0005] To achieve the above purpose, the technical solution of the present invention is: A short-term power load forecasting method based on improved DSSFA-SAC-ConvLSTM, including:

[0006] S1. Collect historical power load data of the electricity consumption load area and perform preprocessing;

[0007] S2. Based on DSSFA analysis, extract features from historical power load data and construct a DSSFA feature matrix;

[0008] S3. Based on the geographical distance of the electricity consumption load area, construct a distance adjacency matrix; based on the DSSFA feature matrix, introduce a self-attention mechanism to generate a dynamic adjacency matrix, and perform adaptive weighted fusion with the distance adjacency matrix, and determine the self-correlation matrix of local space and global space;

[0009] S4. Combine the DSSFA feature matrix, the local spatial autocorrelation matrix, and the global spatial autocorrelation matrix into a splicing matrix, and input it into the ConvLSTM model for feature learning and training to obtain the output feature representation; at the same time, during training, combine the Warmup stage and the improved OneCycleLR stage to adjust the learning rate of the ConvLSTM model;

[0010] S5. Map the output feature representation to obtain the power load prediction value.

[0011] The specific steps of step S2 include:

[0012] S21. Divide the historical power load data into m segments according to time t, and use polynomials for fitting to obtain the trend component, and its expression is as follows:

[0013]

[0014] Where: T(t) is the trend component, t1, t2,..., t m are the segmentation points, and T1(t), T2(t),..., T m (t) are the polynomial fittings for each segment;

[0015] T i (t) = a i，0 + a i，1 t + a i，2 t 2 + … + a i，n t n ;

[0016] Where: T i (t) is the fitting polynomial for the i-th segment, and a i，0 , a i，1 , …, a i，n are the polynomial coefficients for the i-th segment, and n is the order of the polynomial;

[0017] S22. Use a sine function based on dynamic amplitude adjustment to extract the periodic component of the historical power load data, and its expression is as follows:

[0018] P(t) = A(t)·sin(2πft + φ);

[0019] A(t) = A0(+ A1t + A2t 2 ;

[0020] Where: P(t) is the periodic component, A(t) is the dynamically adjusted amplitude, f is the frequency, and φ is the phase;

[0021] S23. Perform singular spectrum analysis on the residual components of historical power load data to obtain long-range correlation components, and its expression is as follows:

[0022]

[0023] Where: R(t) is the long-range correlation component, λ i is the singular value, U i , V i are both singular vectors, T is the transpose symbol, and d is the first d significant components;

[0024] S24. Respectively establish the feature representations of the trend component, periodic component, and long-range correlation component, and integrate them into a matrix to obtain the DSSFA feature matrix, and its expression is as follows:

[0025]

[0026] Where: X DSSFA (t) is the DSSFA feature matrix.

[0027] In the step S23, when performing singular spectrum analysis, it specifically includes:

[0028] S231. For the sequence X = [x1, x2,..., x N of the residual components, construct a trajectory matrix X, and its expression is as follows:

[0029]

[0030] K = N - L + 1;

[0031] Where: K is the number of columns of the trajectory matrix, N is the total length of the sequence, and L is the selected window length;

[0032] S232. Perform singular value decomposition on the trajectory matrix X, and its expression is as follows:

[0033]

[0034] Where: r is the rank of the trajectory matrix;

[0035] S233. Select the first d significant components to reconstruct the noise-free long-range correlation component R(t).

[0036] The step S3 specifically includes:

[0037] S31. Based on the geographical distance between power consumption load regions, construct a distance adjacency matrix, and its expression is as follows:

[0038]

[0039] Where: A distanceis the distance adjacency matrix, d ij is the geographical distance between region i and region j, and β is the attenuation parameter;

[0040] S32. Take the DSSFA feature matrix as the feature matrix and combine it with the self-attention mechanism to generate a dynamic adjacency matrix, and its expression is as follows:

[0041]

[0042] Q = XW Q ; K = XW K ;

[0043] where: A attention is the dynamic adjacency matrix, softmax is the activation function, Q and K are the linear transformations of the feature matrix, and W Q , W K are all learning parameters, and d is the feature dimension;

[0044] S33. Perform adaptive weighted fusion on the distance adjacency matrix and the dynamic adjacency matrix to obtain an adjacency matrix, and its expression is as follows:

[0045] A final = αA distance +(1 - α)A attention ;

[0046]

[0047] where: A final is the adjacency matrix, α is the weight coefficient, ||A distance || F , ||A attention || F are both the Frobenius norms of the adjacency matrix, and a ij is the element in the adjacency matrix;

[0048] S34. Based on the adjacency matrix and the DSSFA feature matrix, calculate the mutual relationship between adjacent regions by weighted calculation, and the expression of the local spatial autocorrelation matrix is as follows:

[0049]

[0050] X i = [T i (t), P i (t), R i (t)];

[0051] X j = [T j (t), P j (t), R j (t)];

[0052]

[0053] Where: L i (t) is the local spatial autocorrelation matrix, X i and X j are the eigenvectors of the DSSFA feature matrix, is the mean of the fused features of all regions, and n is the number of regions;

[0054] S35. Calculate the mutual relationship between all regions by weighted calculation based on the adjacency matrix and the DSSFA feature matrix. Then, the expression of the global spatial autocorrelation matrix is as follows:

[0055]

[0056] Where: G(t) is the global spatial autocorrelation matrix.

[0057] The step S4 specifically includes:

[0058] S41. Broadcast the global spatial autocorrelation matrix to each region and expand it into an N×1 vector. Its expression is as follows:

[0059]

[0060] Where: G broadcasted (t) is the N×1 vector after the broadcast operation;

[0061] S42. Concatenate the DSSFA feature matrix X DSSFA (t), the local spatial autocorrelation matrix L i (t) and G broadcasted (t) to obtain the concatenated matrix X input (t). Its expression is as follows:

[0062] X input (t) = [X DSSFA (t) L(t) G broadcasted (t)];

[0063] S43. Input the concatenated matrix X input (t) into the ConvLSTM model for feature learning training to obtain the output feature representation. Its expression is as follows:

[0064] H t = ConvLSTM(X input (t), H t-1 );

[0065] Where: H tis the output feature representation of the ConvLSTM model at the current time step t, H t-1 is the hidden state information of the previous time step.

[0066] In the Warmup stage, the learning rate is gradually increased at the beginning of training, and its expression is as follows:

[0067]

[0068] where: η t is the learning rate at the t-th step, η start is the initial learning rate in the Warmup stage, t is the current step number, and T warmup is the maximum number of steps in the Warmup stage, is the target learning rate after the end of the Warmup stage;

[0069] In the improved OneCycleLR stage, the training includes an ascending stage, a descending stage, and a cooling stage;

[0070] The expression of the ascending stage is as follows:

[0071]

[0072] where: η base is the initial learning rate in the OneCycleLR stage, is the target learning rate in the OneCycleLR stage, and T up is the number of steps in the ascending stage of OneCycleLR;

[0073] The expression of the descending stage is as follows:

[0074]

[0075] where: η min is the lowest learning rate in the OneCycleLR stage, and T down is the number of steps in the descending stage of OneCycleLR;

[0076] The expression of the cooling stage is as follows:

[0077] η t = η min ·exp(-β(t - T warmup - T up - T down )), t > T warmup + T up + T dowm ;

[0078] where: β is a hyperparameter controlling the cooling rate.

[0079] A short-term electric load forecasting system based on improved DSSFA-SAC-ConvLSTM, the system includes:

[0080] A load data module, used to collect historical electric load data in the power consumption load area and perform preprocessing;

[0081] A DSSFA analysis module, used to extract features from historical electric load data based on DSSFA analysis and construct a DSSFA feature matrix;

[0082] A spatial dependence analysis module, used to construct a distance adjacency matrix based on the geographical distance of the power consumption load area; based on the DSSFA feature matrix, introduce a self-attention mechanism to generate a dynamic adjacency matrix, and perform adaptive weighted fusion with the distance adjacency matrix, and determine the correlation matrix between the local space and the global space;

[0083] A feature learning and training module, used to fuse the DSSFA feature matrix, the local space correlation matrix, and the global space correlation matrix into a splicing matrix, and input it into the ConvLSTM model for feature learning and training to obtain an output feature representation; at the same time during training, combine the Warmup stage and the improved OneCycleLR stage to adjust the learning rate of the ConvLSTM model;

[0084] A load forecasting module, used to map the output feature representation to obtain the electric load forecast value.

[0085] The DSSFA analysis module constructs a DSSFA feature matrix according to the following steps:

[0086] S21. Divide the historical electric load data into m segments according to time t, and use polynomials for fitting to obtain a trend component, and its expression is as follows:

[0087]

[0088] Among them: T(t) is the trend component, t1, t2,..., t m are the segmentation points, T1(t), T2(t),..., T m (t) are the polynomial fittings for each segment;

[0089] T i (t) = a i，0 + a i，1 t + a i，2 t 2 + … + a i，n t n ;

[0090] Among them: T i(t) is the fitting polynomial of the i-th segment, a i,0 , a i,1 ,..., a i,n are the polynomial coefficients of the i-th segment, and n is the order of the polynomial;

[0091] S22. The periodic component of the historical power load data is extracted by using a sine function based on dynamic amplitude adjustment, and its expression is as follows:

[0092] P(t) = A(t)·sin(2πft + φ);

[0093] A(t) = A0 + A1t + A2t 2 ;

[0094] where: P(t) is the periodic component, A(t) is the dynamically adjusted amplitude, f is the frequency, and φ is the phase;

[0095] S23. Singular spectrum analysis is performed on the remaining component of the historical power load data to obtain the long-range correlation component, and its expression is as follows:

[0096]

[0097] where: R(t) is the long-range correlation component, λ i is the singular value, U i , V i are both singular vectors, T is the transpose symbol, and d is the first d significant components;

[0098] S24. The characteristic representations of the trend component, the periodic component, and the long-range correlation component are established respectively and integrated into a matrix to obtain the DSSFA characteristic matrix, and its expression is as follows:

[0099]

[0100] where: X DSSFA (t) is the DSSFA characteristic matrix.

[0101] The spatial dependence analysis module analyzes the regional spatial dependence correlation according to the following steps:

[0102] S31. Based on the geographical distance between power consumption load regions, a distance adjacency matrix is constructed, and its expression is as follows:

[0103]

[0104] where: A distance is the distance adjacency matrix, d ij is the geographical distance between region i and region j, and β is the attenuation parameter;

[0105] S32. Take the DSSFA feature matrix as the feature matrix and combine it with the self-attention mechanism to generate a dynamic adjacency matrix, and its expression is as follows:

[0106]

[0107] Q = XW Q ; K = XW K ;

[0108] where: A attention is the dynamic adjacency matrix, softmax is the activation function, Q and K are the linear transformations of the feature matrix, W Q , W K are both learning parameters, and d is the feature dimension;

[0109] S33. Perform adaptive weighted fusion on the distance adjacency matrix and the dynamic adjacency matrix to obtain an adjacency matrix, and its expression is as follows:

[0110] A final = αA distance +(1 - α)A attention ;

[0111]

[0112] where: A final is the adjacency matrix, α is the weight coefficient, ||A distance || F , ||A attention || F are both the Frobenius norms of the adjacency matrix, and a ij is the element in the adjacency matrix;

[0113] S34. Based on the adjacency matrix and the DSSFA feature matrix, calculate the mutual relationship between adjacent regions by weighted calculation, and the expression of the local spatial autocorrelation matrix is as follows:

[0114]

[0115] X i = [T i (t), P i (t), R i (t)];

[0116] X j = [T j (t), Pj(t), R j (t)];

[0117]

[0118] where: L i(t) is the local spatial autocorrelation matrix, X i and X j are the eigenvectors of the DSSFA feature matrix, is the mean of the fusion features of all regions, and n is the number of regions;

[0119] S35. Calculate the mutual relationship between all regions by weighting the adjacency matrix and the DSSFA feature matrix, then the expression of the global spatial autocorrelation matrix is as follows:

[0120]

[0121] where: G(t) is the global spatial autocorrelation matrix.

[0122] The feature learning and training module adjusts the learning rate according to the following method:

[0123] In the Warmup stage, the learning rate is gradually increased at the beginning of training, and its expression is as follows:

[0124]

[0125] where: η t is the learning rate at the t-th step, η start is the initial learning rate in the Warmup stage, t is the current step number, and T warmup is the maximum step number in the Warmup stage, is the target learning rate after the end of the Warmup stage;

[0126] In the improved OneCycleLR stage, the training includes an ascending stage, a descending stage, and a cooling stage;

[0127] The expression of the ascending stage is as follows:

[0128]

[0129] where: η base is the initial learning rate in the OneCycleLR stage, is the target learning rate in the OneCycleLR stage, and T up is the number of steps in the ascending stage of OneCycleLR;

[0130] The expression of the descending stage is as follows:

[0131]

[0132] where: η min is the lowest learning rate in the OneCycleLR stage, and T down is the number of steps in the descending stage of OneCycleLR;

[0133] The expression in the cooling stage is as follows:

[0134] η t = η min ·exp(-β(t - T warmup - T up - T down ))), t > T warmup + T p + T down ;

[0135] Where: β is a hyperparameter for controlling the cooling rate.

[0136] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0137] In a short-term electric load forecasting method and system based on improved DSSFA-SAC-ConvLSTM of the present invention, the method first collects and preprocesses regional historical electric load data, then uses DSSFA to analyze and extract features to construct a feature matrix. Based on geographical distance and DSSFA features, a distance adjacency matrix and a dynamic adjacency matrix are generated and fused into a self-correlation matrix. Subsequently, the feature matrix and the self-correlation matrix are concatenated and input into the ConvLSTM model for training, and Warmup and improved OneCycleLR are used to adjust the learning rate. Finally, the model output features are mapped to electric load prediction values; in the application of this design, key features are extracted through DSSFA, effectively reducing the data dimension and retaining important information in the data. At the same time, a dynamic adjacency matrix is generated by combining geographical distance and self-attention mechanism, fully considering spatial information, improving the accuracy of multi-region prediction, and using Warmup and improved OneCycleLR strategies to optimize the ConvLSTM learning rate, enhancing the model training effect, further improving the accuracy of electric load prediction, and making it have better performance in complex scenarios. Brief Description of the Drawings

[0138] Figure 1 is the flowchart of the method of the present invention.

[0139] Figure 2 is the comparison chart of prediction results of different methods in Embodiment 1 of the present invention.

[0140] Figure 3 is the system structure diagram of the present invention.

[0141] Figure 4 is the device structure diagram of the present invention.

[0142] In the figure: Load data module 1, DSSFA analysis module 2, spatial dependence analysis module 3, feature learning and training module 4, load forecasting module 5, processor 6, memory 7, computer program code 71. Detailed implementation manner

[0143] The present invention will be further described in detail below with reference to the accompanying drawings and the detailed implementation manner.

[0144] Embodiment 1:

[0145] See Figure 1 , a short-term electric load forecasting method based on improved DSSFA-SAC-ConvLSTM, including:

[0146] S1. Collect historical electric load data of the electricity consumption load area and perform preprocessing;

[0147] In this embodiment, the historical electric load data includes load power data (such as active power, reactive power, apparent power); time series data (such as time stamps, annual, monthly, and daily load curves); user type data (such as residential load, commercial load); geographical area data (such as regional load, regional load curve).

[0148] The preprocessing includes: outlier processing, missing value filling, and data normalization;

[0149] The outlier processing: Identify and correct the outliers in the data through statistical analysis or machine learning methods to avoid the interference of extreme values on the prediction results.

[0150] The missing value filling: Use interpolation method, mean filling or model-based prediction method to complete the missing data to ensure data integrity.

[0151] The data normalization: Standardize or normalize the data to a unified range to eliminate the dimension difference and improve the convergence speed and prediction accuracy of the model.

[0152] S2. Extract features from the historical electric load data based on DSSFA analysis and construct a DSSFA feature matrix;

[0153] In this embodiment, since the electric load data has significant trend, periodicity, and long-range correlation characteristics. The trend reflects the long-term change of the electric load with economic development and seasonal changes, such as the continuous upward trend of the electricity consumption peak in summer and winter; the periodicity shows the fixed regular fluctuations of the daily cycle and the weekly cycle, such as the electricity consumption differences between the morning and evening peaks every day or between weekdays and weekends; the long-range correlation reveals the internal correlation across time scales hidden in the load time series, such as the inertia of the electricity consumption pattern or the continuous influence of the electricity consumption habits in a specific season on the current load.

[0154] These characteristics together constitute the complex time evolution law of the power load, providing a more comprehensive basis for prediction. In response to these characteristics, in this solution, DSSFA analysis is used to decompose, denoise, and reconstruct the power load data, thereby explicitizing the trend, periodicity, and long-range correlation in the data, making the originally complex load signal clearer. Specifically, the DSSFA analysis extracts each component through the following steps: First, a piecewise polynomial is used to extract the trend component to reveal the overall change trend of the load; second, a sine function with dynamically adjusted amplitude is used to extract the periodic component to capture the periodic fluctuation characteristics of the load; finally, singular spectrum analysis (SSA) is applied to the remaining sequence to extract the long-range correlation component to characterize the correlation of the load over a long time span. After each component is denoised, they are recombined to generate a clearer and higher-quality signal. The specific steps are as follows:

[0155] The basic idea of piecewise polynomial fitting is to divide the data into several segments and use polynomials to fit each segment of data to better capture the non-linear trend changes in the data. The specific steps are as follows:

[0156] S21. Divide the historical power load data into m segments according to time t and use polynomials for fitting to obtain the trend component, and its expression is as follows:

[0157]

[0158] Where: T(t) is the trend component; t1, t2,..., t m are the segmentation points, usually the breakpoints and turning points of the data; T1(t), T2(t),..., T m (t) are the polynomial fittings for each segment;

[0159] T i (t) = a i，0 + a i，1 t + a i，2 t 2 + … + a i，n t n ;

[0160] Where: T i (t) is the fitting polynomial for the i-th segment, a i，0 , a i，1 ,..., a i，n are the polynomial coefficients for the i-th segment, and n is the order of the polynomial;

[0161] Although the traditional sine function can capture periodic fluctuations, it has poor adaptability to complex fluctuations. In this solution, the amplitude of the sine function is dynamically adjusted by introducing polynomial correction to adapt to the complexity of the amplitude change in power load data, thereby improving the fitting accuracy of the periodic component. The specific steps are as follows:

[0162] S22. Extract the periodic component of the historical power load data using the sine function with dynamically adjusted amplitude, and its expression is as follows:

[0163] P(t) = A(t)·sin(2πft + φ);

[0164] A(t) = A0 + A1t + A2t 2 ;

[0165] where: P(t) is the periodic component, A(t) is the dynamically adjusted amplitude, f is the frequency, and φ is the phase;

[0166] S23. Perform singular spectrum analysis on the remaining component of the historical power load data to obtain the long-range correlation component, and its expression is as follows:

[0167]

[0168] where: R(t) is the long-range correlation component, λ i is the singular value (reflecting the intensity of the corresponding signal), U i , V i are the singular vectors corresponding to the singular values, T is the transpose symbol, and d is the first d significant components;

[0169] Furthermore, in step S23, when performing singular spectrum analysis, it specifically includes:

[0170] S231. For the sequence X = [x1, x2,..., x N of the remaining component, construct the trajectory matrix X, and its expression is as follows:

[0171]

[0172] K = N - L + 1;

[0173] where: K is the number of columns of the trajectory matrix, N is the total length of the sequence, and L is the selected window length;

[0174] S232. Perform singular value decomposition on the trajectory matrix X, and its expression is as follows:

[0175]

[0176] where: r is the rank of the trajectory matrix;

[0177] S233. Select the first d significant components and reconstruct the noise-free long-range correlation component R(t).

[0178] S24. Respectively establish the characteristic representations of the trend component, the periodic component, and the long-range correlation component. Their expressions are as follows:

[0179] T i (t), i = 1, 2,..., N;

[0180] where: T i (t) is the characteristic representation of the trend component in the i-th region;

[0181] P i (t), i = 1, 2,..., N;

[0182] where: P i (t) is the characteristic representation of the periodic component in the i-th region;

[0183] R i (t), i = 1, 2,..., N;

[0184] where: R i (t) is the characteristic representation of the long-range correlation component in the i-th region;

[0185] And integrate the characteristic representations of the trend component, the periodic component, and the long-range correlation component into a matrix to obtain the DSSFA feature matrix. Its expression is as follows:

[0186]

[0187] where: X DSSFA (t) is the DSSFA feature matrix, which can be used as the input for the subsequent SAC-ConvLSTM model; each row in the matrix corresponds to a region, and each column corresponds to a component extracted by DSSFA.

[0188] S3. Based on the geographical distance of the electricity load regions, construct a distance adjacency matrix; based on the DSSFA feature matrix, introduce a self-attention mechanism to generate a dynamic adjacency matrix, and perform adaptive weighted fusion with the distance adjacency matrix, and determine the self-correlation matrix of the local space and the global space;

[0189] In this embodiment, based on the generated DSSFA feature matrix, an improved SAC module is used to construct spatial dependence features and fuse local and global spatial information. The core of the SAC module (Spatial Auto Correlation, SAC) lies in constructing an adjacency matrix and performing spatial autocorrelation modeling based on it. Therefore, this solution is mainly reflected in the optimization of the adjacency matrix and the spatial autocorrelation modeling method. Spatial autocorrelation refers to the fact that there is a certain degree of correlation in the power load demands of different regions in the geographical or spatial dimension. Specifically, the change in power demand in one region often affects the power demand in adjacent regions, thus showing a dependence relationship or similarity between regions; this characteristic is particularly important in short-term power load forecasting. By introducing spatial autocorrelation, the dynamic changes in load between regions can be captured more accurately, improving the accuracy and reliability of forecasting.

[0190] The key to spatial autocorrelation lies in the adjacency matrix A, which is used to represent the spatial relationship between regions. In short-term power load forecasting, the element A in the proximity matrix A ij reflects the spatial correlation or influence of the power load in region i on region j. However, the traditional adjacency matrix lacks flexibility and adaptability and is difficult to model complex dynamic spatial dependence relationships. For this reason, in this solution, a self-attention mechanism is introduced to generate a dynamic adjacency matrix, which is adaptively weighted and fused with the distance adjacency matrix, taking into account both feature-driven dynamic associations and fixed physical topologies. The fused combined adjacency matrix can not only dynamically capture the spatial characteristics of power loads but also maintain the robustness of physical spatial relationships, thus more comprehensively capturing local and global spatial dependencies and enhancing the model's ability to express complex spatial relationships, which helps to more accurately perform power load forecasting. The specific steps are as follows:

[0191] S31. Since the power loads in regions that are closer may affect each other, a distance adjacency matrix can be constructed based on the geographical distance between electricity consumption load regions, and its expression is as follows:

[0192]

[0193] where: A distance is the distance adjacency matrix; d ij is the geographical distance between region i and region j; β is the attenuation parameter used to control the influence of distance;

[0194] S32. Using the DSSFA feature matrix as the feature matrix and combining it with the self-attention mechanism to generate a dynamic adjacency matrix, and its expression is as follows:

[0195]

[0196] Q = XW Q ; K = XWK ;

[0197] Where: A attention is the dynamic adjacency matrix; softmax is the activation function; Q and K are linear transformations of the feature matrix; W Q , W K are all learning parameters; d is the feature dimension, which is used to prevent numerical explosion;

[0198] S33. Fuse the distance adjacency matrix and the dynamic adjacency matrix through adaptive weighting to obtain the adjacency matrix, and its expression is as follows:

[0199] A final = αA distance + (1 - α)A attention ;

[0200]

[0201] Where: A final is the adjacency matrix; α is the weight coefficient, which is used to adjust the ratio of attention to the distance adjacency matrix; ||A distance || F , ||A attention || F are both the Frobenius norms of the adjacency matrix, which are used to measure the energy of the matrix; a ij is an element in the adjacency matrix; The meaning of is: if the information content of A distance increases, then increase α, if the information content of A distance decreases, then decrease α, so as to achieve adaptive adjustment.

[0202] After calculating the adjacency matrix, it is necessary to calculate the local spatial autocorrelation and the global spatial autocorrelation. The global spatial autocorrelation measures the overall load dependence between regions through the adjacency matrix, reveals the global spatial pattern of the power load, and helps to identify the correlation between regions; while the local spatial autocorrelation calculates the relationship between each region and its neighboring regions through the adjacency matrix, which helps to discover local hotspots or cold spots and further improve the prediction ability of the load fluctuation in a specific region; the specific steps are as follows:

[0203] S34. Calculate the mutual relationship between adjacent regions based on the weighted adjacency matrix and the DSSFA feature matrix, then the expression of the local spatial autocorrelation matrix is as follows:

[0204]

[0205] X i = [T i (t), P i (t), R i (t)];

[0206] X j = [T j (t), P j (t), R j (t)];

[0207]

[0208] Where: L i (t) is the local spatial autocorrelation matrix; X i , X j are the eigenvectors of the DSSFA feature matrix, that is, the eigenvectors after the fusion of regions i and j, containing the comprehensive information of the trend, period, and long-range correlation components; is the mean value of the fusion features of all regions; n is the number of regions;

[0209] In this solution, the global spatial autocorrelation is usually measured by Moran's I (Moran index), and the mutual relationship between all regions can be weighted and calculated by using the combined adjacency matrix A final and X DSSFA (t). The specific steps are as follows:

[0210] S35. Based on the adjacency matrix and the DSSFA feature matrix, the mutual relationship between all regions is weighted and calculated. Then the expression of the global spatial autocorrelation matrix is as follows:

[0211]

[0212] Where: G(t) is the global spatial autocorrelation matrix.

[0213] S4. The DSSFA feature matrix, the local spatial autocorrelation matrix, and the global spatial autocorrelation matrix are fused into a splicing matrix and input into the ConvLSTM model for feature learning and training to obtain the output feature representation; at the same time, during training, the Warmup stage and the improved OneCycleLR stage are combined to adjust the learning rate of the ConvLSTM model;

[0214] In this embodiment, the feature matrix output by DSSFA and the spatial features extracted by SAC are spliced and input into ConvLSTM to form a deep understanding of the spatio-temporal dynamics, and an improved multi-stage dynamic learning rate adjustment strategy is designed to improve the model optimization efficiency. The specific steps are as follows:

[0215] X DSSFA (t) is an N×3 matrix, and the local spatial autocorrelation L i(t) is an N×1 matrix, and G(t) is the global spatial autocorrelation value, which is a scalar with a dimension of 1. To concatenate with other feature matrices, G(t) needs to be broadcast to each region first and extended to an N×1 vector. The specific steps are as follows:

[0216] S41. Broadcast the global spatial autocorrelation matrix to each region and extend it to an N×1 vector. The expression is as follows:

[0217]

[0218] where: G broadcasted (t) is the N×1 vector after the broadcast operation;

[0219] S42. Concatenate the DSSFA feature matrix X DSSFA (t), the local spatial autocorrelation matrix L i (t) and G broadcasted (t) to obtain the concatenated matrix X input (t). The expression is as follows:

[0220] X input (t) = [X DSSFA (t) L(t) G broadcasted (t)];

[0221] ConvLSTM is a model that combines a convolutional neural network (CNN) and a long short-term memory network (LSTM), which can process time series data while capturing spatial features. And X input (t) is obtained by concatenating X DSSFA (t), L i (t) and G broadcasted (t), which already contains certain time characteristics and spatial characteristics. After inputting X input (t) into ConvLSTM, the model can further perform in-depth modeling on time characteristics and spatial dependencies, thereby mining more complex feature relationships, enhancing the ability to capture the law of power load, and achieving a more accurate prediction effect.

[0222] S43. Input the concatenated matrix X input (t) into the ConvLSTM model for feature learning and training to obtain the output feature representation. The expression is as follows:

[0223] H t = ConvLSTM(X input (t), H t-1 );

[0224] where: H tis the output feature representation of the ConvLSTM model at the current time step t, which captures the current input feature X input (t) and the hidden state information H at previous time steps t-1 , aiming to capture the high-dimensional feature information of the current power load state and provide more reliable and rich inputs for the next load prediction. The output H t can be used as the input of the fully connected layer to generate the final load prediction result y(t).

[0225] In this embodiment, after each training stage, the loss function is used to calculate the error between the predicted value and the true value. When the loss function tends to be stable or reaches the preset maximum number of training epochs, the training is stopped to obtain the optimized model, and the optimized model is applied to the actual power load prediction task to predict and evaluate the unseen data.

[0226] Furthermore, in the ConvLSTM model, the weights and biases are the core parameters learned by the model, which determine the calculation method of the network when processing input data. The weights connect the layers and determine the influence of each input signal on the output, while the biases help the model adapt to the offset of the input data. The change of the learning rate will directly affect the update amplitude of the weights and biases, thus affecting the parameter optimization of the convolutional neural network (CNN) and the long short-term memory network (LSTM) in the ConvLSTM model during the training process; a larger learning rate may lead to too large an update step and unstable training; a smaller learning rate may lead to too slow a convergence speed and affect the optimization effect of the model. Therefore, an appropriate learning rate is crucial for the effectiveness and efficiency of the optimization process.

[0227] During the training process of the ConvLSTM model, the weight parameters are updated through backpropagation and gradient descent, and its expression is as follows:

[0228]

[0229] where: W is the weight, b is the bias, is the loss function, and are the gradients of the weight and the bias respectively, and η is the learning rate. However, the traditional learning rate adjustment method lacks flexibility and is difficult to adapt dynamically according to different stages of the training process, which easily leads to unstable convergence in the initial stage, getting stuck in local optima, or insufficient training efficiency and performance. Therefore, in this solution, the Warmup stage and the improved OneCycleLR stage are combined to provide a multi-stage dynamic learning rate adjustment strategy for the ConvLSTM model, which can not only stabilize the initial training but also avoid local optima through periodic changes and improve the performance of the model in power load prediction.

[0230] In the Warmup stage, the model gradually increases the learning rate at the beginning of training to help the model stably enter the convergence stage. Its expression is as follows:

[0231]

[0232] Where: η t is the learning rate at the t-th step, η start is the initial learning rate in the Warmup stage, t is the current step number, and T warmup is the maximum number of steps in the Warmup stage, is the target learning rate after the end of the Warmup stage;

[0233] The traditional OneCycleLR stage only includes the ascending stage and the descending stage. In high-precision prediction tasks such as power load prediction, the traditional descending stage may not be sufficient to fully optimize the model. Therefore, in this solution, after the descending stage of OneCycleLR, a cooling stage is introduced to gradually reduce the learning rate to a very small value. This additional stage helps to further fine-tune the model's parameters in the last few rounds of training, thus ensuring that the model can achieve the best performance, avoid overfitting, and improve the generalization ability.

[0234] The improved OneCycleLR stage includes an ascending stage, a descending stage, and a cooling stage during training;

[0235] The expression of the ascending stage is as follows:

[0236]

[0237] Where: η base is the initial learning rate of the OneCycleLR stage, is the target learning rate of the OneCycleLR stage, and T up is the number of steps in the ascending stage of OneCycleLR;

[0238] The expression of the descending stage is as follows:

[0239]

[0240] Where: η min is the lowest learning rate of the OneCycleLR stage, and T down is the number of steps in the descending stage of OneCycleLR;

[0241] The expression of the cooling stage is as follows:

[0242] η t = η min ·exp(-β(t - T warmup-T up -T down )) when t > T warmup +T up +T down ;

[0243] where: β is a hyperparameter for controlling the cooling rate.

[0244] S5. Map the output feature representation to obtain the power load prediction value.

[0245] The final load prediction result is a series of power load data points, which form a load prediction curve; the core of power load prediction is to use historical load data to predict the future trend of power load changes. Specifically, by inputting the feature H output by the ConvLSTM network t into the fully connected layer to generate the final load prediction result curve; subsequently, compare the generated prediction curve with the true load data curve in the validation set to evaluate the accuracy of the prediction model.

[0246] See Figure 2 , in this embodiment, the prediction results of this solution are compared with those of other different methods. It can be seen from the figure that based on the improved DSSFA-SAC-ConvLSTM model proposed in this solution, its prediction curve is closest to the original load curve, showing higher accuracy in capturing load data features. This model can not only maintain a high prediction accuracy when the load changes smoothly, but also effectively extract feature information during the stage of large load fluctuations. In contrast, the SAC-ConvLSTM and ConvLSTM models have certain deviations when the load changes violently, the gap between the prediction results and the actual values is large, and there are easy to appear lag or advance phenomena, resulting in a decrease in the model accuracy. Through comparison, it can be seen that the improved DSSFA-SAC-ConvLSTM model performs better in the load prediction task and has stronger adaptability in dealing with complex spatio-temporal sequence data.

[0247] The following table shows the comparison of performance indicators of different prediction models, where MAPE (Mean Absolute Percentage Error) and RMSE (Root Mean Squared Error). MAPE reflects the relative error between the predicted value and the actual value, while RMSE emphasizes the absolute magnitude of the prediction error, especially the influence of larger errors.

[0248] Prediction model MAPE(%) RMSE / KW ConvLSTM 2.23 237.36 SAC-ConvLSTM 1.67 176.49 Improved DSSFA-SAC-ConvLSTM proposed in this scheme 1.18 99.78

[0249] From the data in the above table, it can be seen that the improved DSSFA-SAC-ConvLSTM proposed in this solution is significantly superior to the other two models in terms of performance. Specifically, the MAPE of ConvLSTM is 2.23%, SAC-ConvLSTM reduces it to 1.67%, and the improved DSSFA-SAC-ConvLSTM is further optimized to 1.19%, significantly improving the prediction accuracy; in terms of the RMSE index, ConvLSTM is 237.36 kW, SAC-ConvLSTM drops to 176.49 kW, and the improved DSSFA-SAC-ConvLSTM is optimized to 99.85 kW, with a more prominent reduction in error. In summary, DSSFA-SAC-ConvLSTM shows significant advantages in both MAPE and RMSE indicators, especially the improvement in RMSE, which reflects its higher potential in practical load forecasting applications.

[0250] Example 2:

[0251] See Figure 3 , a short-term power load forecasting system based on improved DSSFA-SAC-ConvLSTM, the system includes:

[0252] Load data module 1, used to collect historical power load data of the power consumption load area and perform preprocessing;

[0253] DSSFA analysis module 2, used to extract features from historical power load data based on DSSFA analysis and construct a DSSFA feature matrix;

[0254] Further, the DSSFA analysis module 2 constructs a DSSFA feature matrix according to the following steps:

[0255] S21. Divide the historical power load data into m segments according to time t, and use polynomials for fitting to obtain the trend component, and its expression is as follows:

[0256]

[0257] Among them: T(t) is the trend component, t1, t2,..., t m is the segmentation point, T1(t), T2(t),..., T m (t) is the polynomial fitting of each segment;

[0258] T i (t) = a i，0 + a i，1 t + a i，2 t 2 + … + a i，n t n ;

[0259] Where: T i (t) is the fitting polynomial of the i-th segment, a i,0 , a i,1 ,..., a i,n are the polynomial coefficients of the i-th segment, and n is the order of the polynomial;

[0260] S22. The periodic component of the historical power load data is extracted by using a sine function based on dynamic amplitude adjustment, and its expression is as follows:

[0261] P(t) = A(t)·sin(2πft + φ);

[0262] A(t) = A0 + A1t + A2t 2 ;

[0263] Where: P(t) is the periodic component, A(t) is the dynamically adjusted amplitude, f is the frequency, and φ is the phase;

[0264] S23. Singular spectrum analysis is performed on the remaining component of the historical power load data to obtain the long-range correlation component, and its expression is as follows:

[0265]

[0266] Where: R(t) is the long-range correlation component, λ i is the singular value, U i , V i are both singular vectors, T is the transpose symbol, and d is the first d significant components;

[0267] S24. Feature representations of the trend component, periodic component, and long-range correlation component are established respectively, integrated into a matrix, and the DSSFA feature matrix is obtained, and its expression is as follows:

[0268]

[0269] Where: X DSSFA (t) is the DSSFA feature matrix.

[0270] The spatial dependence analysis module 3 is used to construct a distance adjacency matrix based on the geographical distance of the electricity consumption load area; based on the DSSFA feature matrix, a self-attention mechanism is introduced to generate a dynamic adjacency matrix, and it is adaptively weighted and fused with the distance adjacency matrix, and the correlation matrix of the local space and the global space is determined;

[0271] Furthermore, the spatial dependence analysis module 3 analyzes the regional spatial dependence correlation according to the following steps:

[0272] S31. Based on the geographical distance between electricity consumption load areas, a distance adjacency matrix is constructed, and its expression is as follows:

[0273]

[0274] Among them: A distance is the distance adjacency matrix, d ij is the geographical distance between region i and region j, and β is the attenuation parameter;

[0275] S32. Take the DSSFA feature matrix as the feature matrix and combine it with the self-attention mechanism to generate a dynamic adjacency matrix, and its expression is as follows:

[0276]

[0277] Q = XW Q ; K = XW K ;

[0278] Among them: A attention is the dynamic adjacency matrix, softmax is the activation function, Q and K are the linear transformations of the feature matrix, and W Q , W K are both learning parameters, and d is the feature dimension;

[0279] S33. Adaptively weight and fuse the distance adjacency matrix and the dynamic adjacency matrix to obtain an adjacency matrix, and its expression is as follows:

[0280] A final = αA distance +(1 - α)A attention ;

[0281]

[0282] Among them: A final is the adjacency matrix, α is the weight coefficient, ||A distance || F , ||A attention || F are both the Frobenius norms of the adjacency matrix, and a ij is the element in the adjacency matrix;

[0283] S34. Based on the adjacency matrix and the DSSFA feature matrix, calculate the mutual relationship between adjacent regions by weighted calculation, and the expression of the local spatial autocorrelation matrix is as follows:

[0284]

[0285] X i = [T i (t), P i (t), R i (t)];

[0286] Xj = [T j (t), P j (t), R j (t)];

[0287]

[0288] Where: L i (t) is the local spatial autocorrelation matrix, X i , X j are the eigenvectors of the DSSFA feature matrix, is the mean of the fused features of all regions, and n is the number of regions;

[0289] S35. Calculate the mutual relationship between all regions by weighted calculation of the adjacency matrix and the DSSFA feature matrix. Then the expression of the global spatial autocorrelation matrix is as follows:

[0290]

[0291] Where: G(t) is the global spatial autocorrelation matrix.

[0292] The feature learning and training module 4 is used to fuse the DSSFA feature matrix, the local spatial correlation matrix, and the global spatial correlation matrix into a concatenated matrix, and input it into the ConvLSTM model for feature learning and training to obtain the output feature representation; at the same time, during training, the Warmup stage and the improved OneCycleLR stage are combined to adjust the learning rate strategy of the ConvLSTM model;

[0293] Furthermore, the feature learning and training module 4 adjusts the learning rate strategy according to the following method:

[0294] In the Warmup stage, the learning rate is gradually increased at the beginning of training, and its expression is as follows:

[0295]

[0296] Where: η t is the learning rate at the t-th step, η start is the initial learning rate in the Warmup stage, t is the current step number, T warmup is the maximum number of steps in the Warmup stage, is the target learning rate after the end of the Warmup stage;

[0297] In the improved OneCycleLR stage, it includes an ascending stage, a descending stage, and a cooling stage during training;

[0298] The expression of the ascending stage is as follows:

[0299]

[0300] where: η base is the initial learning rate in the OneCycleLR stage, is the target learning rate in the OneCycleLR stage, T up is the number of steps in the rising stage of OneCycleLR;

[0301] The expression for the descending stage is as follows:

[0302]

[0303] where: η min is the minimum learning rate in the OneCycleLR stage, T down is the number of steps in the descending stage of OneCycleLR;

[0304] The expression for the cooling stage is as follows:

[0305] η t = η min ·exp(-β(t - T warmup - T up - T down )), t > T warmup + T up + T down ;

[0306] where: β is a hyperparameter for controlling the cooling rate.

[0307] The load prediction module 5 is used to map the output feature representation to obtain the power load prediction value.

[0308] Example 3:

[0309] See Figure 4 , a short-term power load prediction device based on improved DSSFA-SAC-ConvLSTM, the device includes a processor 6 and a memory 7;

[0310] The memory 7 is used to store the computer program code 71 and transmit the computer program code 71 to the processor 6;

[0311] The processor 6 is used to execute the short-term power load prediction method based on improved DSSFA-SAC-ConvLSTM described in Example 1 according to the instructions in the computer program code 71.

[0312] In this embodiment, there is also provided a computer-readable storage medium storing computer-executable instructions, which, when executed on a computer, implement the short-term electric load forecasting method based on the improved DSSFA-SAC-ConvLSTM described in Embodiment 1.

[0313] Generally speaking, the computer instructions for implementing the method of the present invention can be carried by any combination of one or more computer-readable storage media. A non-transitory computer-readable storage medium can include any computer-readable medium except for the signals propagating temporarily per se.

[0314] The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples (non-exhaustive list) of the computer-readable storage medium include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0315] The computer program code for performing the operations of the present invention can be written in one or more programming languages or combinations thereof. The programming languages include object-oriented programming languages such as Java, Smalltalk, C++, and also include conventional procedural programming languages such as the "C" language or similar programming languages. In particular, the Python language suitable for neural network computing and platform frameworks based on TensorFlow, PyTorch, etc. can be used. The program code can be executed entirely on the user's computer, partially on the user's computer, executed as an independent software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or connected to an external computer (for example, connected through the Internet using an Internet service provider).

[0316] For the above-mentioned device and non-transitory computer-readable storage medium, reference can be made to the specific description of a short-term electric load forecasting method based on the improved DSSFA-SAC-ConvLSTM and its beneficial effects, which will not be elaborated here.

[0317] Although the embodiments of the present invention have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A short-term electric load forecasting method based on improved DSSFA-SAC-ConvLSTM, characterized in that, Including: S1. Collect historical power load data of the power consumption load area and perform preprocessing; S2. Extract features from the historical power load data based on DSSFA analysis and construct a DSSFA feature matrix; S3. Based on the geographical distance between power consumption load areas, construct a distance adjacency matrix; based on the DSSFA feature matrix, introduce a self-attention mechanism to generate a dynamic adjacency matrix, and perform adaptive weighted fusion with the distance adjacency matrix, and determine the self-correlation matrix of the local space and the global space; S4. Fuse the DSSFA feature matrix, the local space self-correlation matrix, and the global space self-correlation matrix into a concatenated matrix and input it into the ConvLSTM model for feature learning training to obtain an output feature representation; at the same time during training, combine the Warmup stage and the improved OneCycleLR stage to adjust the learning rate of the ConvLSTM model; S5. Map the output feature representation to obtain a power load prediction value.

2. The short-term power load prediction method based on improved DSSFA-SAC-ConvLSTM according to claim 1, wherein: The step S2 specifically includes: S21. Divide the historical power load data into m segments according to time t and perform fitting using a polynomial to obtain a trend component, and its expression is as follows: Where: T(t) is the trend component, t1, t2, ..., t m are the segmentation points, and T1(t), T2(t), ..., T m (t) are the polynomial fittings for each segment; T i (t) = a i，0 + a i，1 t + a i，2 t 2 + … + a i，n t n ; where: T i (t) is the fitting polynomial of the i-th segment, a i，0 , a i，1 ,..., a i，n are the polynomial coefficients of the i-th segment, and n is the order of the polynomial; S22. Extract the periodic component of the historical power load data using a sine function based on dynamic amplitude adjustment, and its expression is as follows: P(t) = A(t)·sin(2πft + φ); A(t) = A0 + A1t + A2t 2 ; where: P(t) is the periodic component, A(t) is the dynamically adjusted amplitude, f is the frequency, and φ is the phase; S23. Perform singular spectrum analysis on the remaining component of the historical power load data to obtain a long-range correlation component, and its expression is as follows: Among them: R(t) is the long-range correlation component, λ i is the singular value, U i , V i are both singular vectors, T is the transpose symbol, and d is the first d significant components; S24. Respectively establish feature representations of the trend component, the periodic component, and the long-range correlation component, and integrate them into a matrix to obtain a DSSFA feature matrix, and its expression is as follows: Where: X DSSFA (t) is the DSSFA feature matrix.

3. The short-term power load prediction method based on improved DSSFA-SAC-ConvLSTM according to claim 2, wherein: In the step S23, when performing singular spectrum analysis, it specifically includes: S231. For the sequence X = [x1, x2,..., x N , construct a trajectory matrix X, and its expression is as follows: K = N - L + 1; where: K is the number of columns of the trajectory matrix, N is the total length of the sequence, and L is the selected window length; S232. Perform singular value decomposition on the trajectory matrix X, and its expression is as follows: where: r is the rank of the trajectory matrix; S233. Select the first d significant components to reconstruct the noise-free long-range correlation component R(t).

4. The short-term power load prediction method based on improved DSSFA-SAC-ConvLSTM according to claim 1, wherein: The step S3 specifically includes: S31. Based on the geographical distance between power consumption load areas, construct a distance adjacency matrix, and its expression is as follows: Where: A distance is the distance adjacency matrix, d ij is the geographical distance between region i and region j, and β is the attenuation parameter; S32. Use the DSSFA feature matrix as the feature matrix and combine it with the self-attention mechanism to generate a dynamic adjacency matrix, and its expression is as follows: Q = XW Q ; K = XW K ; Where: A attention is a dynamic adjacency matrix, softmax is an activation function, Q and K are linear transformations of the feature matrix, and W Q , W K are all learning parameters, and d is the feature dimension; S33. Perform adaptive weighted fusion on the distance adjacency matrix and the dynamic adjacency matrix to obtain an adjacency matrix, and its expression is as follows: A final = αA distance + (1 - α)A attention ; Among them: A final is the adjacency matrix, α is the weight coefficient, ||A distance || F and ‖A attention || F are both the Frobenius norms of the adjacency matrix, a ij is an element in the adjacency matrix; S34. Calculate the mutual relationship between adjacent regions based on the weighted adjacency matrix and DSSFA feature matrix. The expression of the local spatial autocorrelation matrix is as follows: X i = [T i (t), P i (t), R i (t)]; X j = [T j (t), P j (t), R j (t)]; Where: L i (t) is the local spatial autocorrelation matrix, X i , X j are the eigenvectors of the DSSFA feature matrix, is the mean of the fusion features of all regions, and n is the number of regions; S35. Calculate the mutual relationship between all regions based on the weighted adjacency matrix and DSSFA feature matrix. The expression of the global spatial autocorrelation matrix is as follows: Where: G(t) is the global spatial autocorrelation matrix.

5. The short-term power load forecasting method based on improved DSSFA-SAC-ConvLSTM according to claim 1, characterized in that: The step S4 specifically includes: S41. Broadcast the global spatial autocorrelation matrix to each region and expand it into a vector of N×1. The expression is as follows: Where: G broadcasted (t) is an N×1 vector after the broadcast operation; S42. Concatenate the DSSFA feature matrix X DSSFA (t), the local spatial autocorrelation matrix L i (t) and G broadcasted (t) to obtain the concatenated matrix X input (t), and its expression is as follows: X inpu (t) = [X DSSFA (t) L(t) G broadcasted (0)]; S43. Input the splicing matrix X input (t) into the ConvLSTM model for feature learning training to obtain the output feature representation, and its expression is as follows: H t = ConvLSTM(X input (t), H t-1 ); Where: H t is the output feature representation of the ConvLSTM model at the current time step t, and H t-1 is the hidden state information of the previous time step.

6. The short-term power load forecasting method based on improved DSSFA-SAC-ConvLSTM according to claim 1, characterized in that: In the Warmup stage, gradually increase the learning rate at the initial stage of training. The expression is as follows: where: η t is the learning rate at the t-th step, η start is the initial learning rate in the Warmup stage, t is the current step number, T warmup is the maximum number of steps in the Warmup stage, is the target learning rate after the end of the Warmup stage; In the improved OneCycleLR stage, it includes an ascending stage, a descending stage, and a cooling stage during training; The expression of the ascending stage is as follows: Where: η base is the initial learning rate in the OneCycleLR stage, is the target learning rate in the OneCycleLR stage, T up is the number of steps in the rising stage of OneCycleLR; The expression of the descending stage is as follows: Where: η min is the lowest learning rate in the OneCycleLR stage, and T down is the number of steps in the OneCycleLR decay stage; The expression of the cooling stage is as follows: η t = η min · exp(-β(t - T warmup - T up - T down ))), t > T warmup + T up + T down ; Where: β is a hyperparameter controlling the cooling rate.

7. A short-term electric load forecasting system based on improved DSSFA-SAC-ConvLSTM, characterized in that, The system includes: A load data module (1) for collecting historical power load data of the power consumption load area and performing preprocessing; A DSSFA analysis module (2) for extracting features from historical power load data based on DSSFA analysis and constructing a DSSFA feature matrix; A spatial dependence analysis module (3) for constructing a distance adjacency matrix based on the geographical distance of the power consumption load area; introducing a self-attention mechanism based on the DSSFA feature matrix to generate a dynamic adjacency matrix, and performing adaptive weighted fusion with the distance adjacency matrix, and determining the correlation matrix of the local space and the global space; A feature learning and training module (4) for fusing the DSSFA feature matrix, the local spatial correlation matrix, and the global spatial correlation matrix into a splicing matrix and inputting it into a ConvLSTM model for feature learning and training to obtain an output feature representation; at the same time during training, combining the Warmup stage and the improved OneCycleLR stage to adjust the learning rate of the ConvLSTM model; A load forecasting module (5) for mapping the output feature representation to obtain a power load forecasting value.

8. The short-term power load forecasting system based on improved DSSFA-SAC-ConvLSTM according to claim 7, characterized in that: The DSSFA analysis module (2) constructs a DSSFA feature matrix according to the following steps: S2 1. Divide the historical power load data into m segments according to time t and perform fitting with a polynomial to obtain a trend component. The expression is as follows: where: T(t) is the trend component, t1, t2, ..., t m are the segmentation points, and T1(t), T2(t), ..., T m (t) are the polynomial fittings for each segment; T i (t) = a i，0 + a i，1 t + a i，2 t 2 + … + a i，n t n ; Where: T i (t) is the fitting polynomial of the i-th segment, a i，0 , a i，1 ,..., a i，n are the polynomial coefficients of the i-th segment, and n is the order of the polynomial; S22. Extract the periodic component of the historical power load data using a sine function based on dynamic amplitude adjustment. The expression is as follows: P(t) = A(t)·sin(2πft + φ); A(t) = A0 + A1t + A2t 2 ; Where: P(t) is the periodic component, A(t) is the dynamically adjusted amplitude, f is the frequency, and φ is the phase; S23. Perform singular spectrum analysis on the residual component of historical power load data to obtain the long-range correlation component, and its expression is as follows: Among them: R(t) is the long-range correlation component, λ i is the singular value, U i , V i are both singular vectors, T is the transpose symbol, and d is the first d significant components; S24. Respectively establish the feature representations of the trend component, periodic component, and long-range correlation component, integrate them into a matrix, and obtain the DSSFA feature matrix, and its expression is as follows: Where: X DSSFA (t) is the DSSFA feature matrix.

9. The short-term power load forecasting system based on improved DSSFA-SAC-ConvLSTM according to claim 7, characterized in that: The spatial dependence analysis module (3) analyzes the regional spatial dependence correlation according to the following steps: S31. Based on the geographical distance between power consumption load areas, construct a distance adjacency matrix, and its expression is as follows: Among them: A distance is the distance adjacency matrix, d ij is the geographical distance between region i and region j, and β is the attenuation parameter; S32. Use the DSSFA feature matrix as the feature matrix and combine it with the self-attention mechanism to generate a dynamic adjacency matrix, and its expression is as follows: Q = XW Q ; KK = XW K ; Where: A attention is a dynamic adjacency matrix, softmax is an activation function, Q and K are linear transformations of the feature matrix, and W Q , W K are all learning parameters, and d is the feature dimension; S33. Perform adaptive weighted fusion on the distance adjacency matrix and the dynamic adjacency matrix to obtain the adjacency matrix, and its expression is as follows: A final = αA distance + (1 - α)A attention ; Where: A final is the adjacency matrix, α is the weight coefficient, ||A distance || F and ||A attention || F are both the Frobenius norms of the adjacency matrix, and a ij is an element in the adjacency matrix; S34. Based on the weighted calculation of the adjacency matrix and the DSSFA feature matrix for the mutual relationship between adjacent regions, the local spatial autocorrelation matrix expression is as follows: X i = [T i (t), P i (t), R i (t)]; X j = [[T j (t), P j (t), R j (t)]; Where: L i (t) is the local spatial autocorrelation matrix, X i , X j is the eigenvector of the DSSFA feature matrix, is the mean of the fusion features of all regions, and n is the number of regions; S35. Based on the weighted calculation of the adjacency matrix and the DSSFA feature matrix for the mutual relationship between all regions, the global spatial autocorrelation matrix expression is as follows: Where: G(t) is the global spatial autocorrelation matrix.

10. The short-term power load forecasting system based on improved DSSFA-SAC-ConvLSTM according to claim 7, characterized in that: The feature learning and training module (4) adjusts the learning rate strategy in the following manner: In the Warmup stage, gradually increase the learning rate at the initial stage of training, and its expression is as follows: where: η t is the learning rate at the t-th step, η start is the initial learning rate in the Warmup stage, t is the current step number, T warmup is the maximum number of steps in the Warmup stage, is the target learning rate after the end of the Warmup stage; In the improved OneCycleLR stage, the training includes an ascending stage, a descending stage, and a cooling stage; The expression of the ascending stage is as follows: Where: η base is the initial learning rate in the OneCycleLR stage, is the target learning rate in the OneCycleLR stage, T up is the number of steps in the rising stage of OneCycleLR; The expression of the descending stage is as follows: Where: η min is the lowest learning rate in the OneCycleLR stage, and T down is the number of steps in the OneCycleLR decay stage; The expression of the cooling stage is as follows: η t = η min · exp(-β(t - T warmup - T up - T down ))), t > T warmup + T up + T down ; Where: β is the hyperparameter controlling the cooling speed.

Images