A short-term load forecasting method and system based on trend and seasonal modeling

Abstract

The application discloses a short-term load prediction method and system based on trend and seasonal modeling, and belongs to the technical field of electric power load prediction. The method comprises the following steps: obtaining to-be-tested load data; pre-processing the to-be-tested load data to obtain standard load data; and inputting the standard load data into a pre-trained short-term load prediction model to obtain a short-term load prediction value. The method for obtaining the pre-trained short-term load prediction model comprises the following steps: obtaining a historical load data set and pre-processing each historical load data in the historical load data set to obtain a historical sample set; and inputting the historical sample set into a pre-constructed short-term load prediction model to obtain a trained short-term load prediction model. The pre-constructed short-term load prediction model is based on a weight self-adaptive adjustment linear layer to extract trend characteristics and based on an extended long short-term memory network to extract seasonal characteristics, so that the effect of high-precision load prediction of load data in a short time scale is achieved.

Description

Technical Field

[0001] This invention belongs to the field of power load forecasting technology, and particularly relates to a short-term load forecasting method and system based on trend and seasonal modeling. Background Technology

[0002] In modern power systems, load forecasting is a crucial element in ensuring the safe, economical, and efficient operation of power supply. With the rapid growth of global energy demand, accurate load forecasting plays a vital role in enabling power companies to optimize generation plans, rationally allocate resources, and effectively respond to load fluctuations. Especially against the backdrop of an increasing proportion of renewable energy, the volatility and uncertainty of power load have further intensified, posing new challenges to traditional load forecasting methods.

[0003] However, due to the complexity of time series forecasting and the diversity of application scenarios, using traditional models for forecasting can be extremely difficult. Therefore, in recent years, modern methods such as Artificial Neural Networks (ANNs), Random Forests, and Support Vector Machines (SVMs) have gradually become research hotspots in power load forecasting. Artificial Neural Networks, as a powerful nonlinear modeling tool, can capture complex nonlinear relationships, but the training process requires a large amount of data and computational resources. With the rapid development of deep learning technology, various neural network models have been introduced into time series forecasting tasks. Among them, Transformer and its variants (such as Reformer and Temporal Fusion Transformer) have attracted much attention due to their powerful self-attention mechanism. Transformer significantly improves training speed through parallel processing, while Reformer optimizes the modeling efficiency of long sequences by introducing Locality Sensitive Hashing (LSH) and Reversible Residual Networks. However, the high computational complexity and resource requirements of these models limit their application in scenarios with high real-time requirements. Convolutional Neural Networks (CNNs) have also achieved significant results in time series forecasting, especially in capturing local features. 1D convolutional networks are effective at extracting short-term patterns in time series data, but they typically perform poorly when dealing with long-term dependencies and complex seasonal variations. Furthermore, Long Short-Term Memory (LSTM) networks and their variants (such as Gated Recurrent Units, GRUs) are gaining popularity due to their built-in gating mechanism, which effectively mitigates the vanishing gradient problem, making them suitable for modeling long-term dependencies. However, LSTM networks have high computational complexity, especially when dealing with extremely large datasets, and training speeds are often slow. Random forests are a tree-based ensemble learning method that improves prediction accuracy by constructing multiple decision trees. While they perform well with high-dimensional features, their model complexity and training time increase with time series data, making them less efficient for real-time prediction. Support Vector Machines (SVMs) excel in few-shot learning and can effectively handle complex data through nonlinear kernel functions, but they may face computational burdens and overfitting issues when dealing with large-scale electricity load data, especially when the data volume is large and the feature dimensions are high.

[0004] While the aforementioned modern methods have improved prediction accuracy to some extent, they still suffer from high model complexity, long training times, and high computational resource consumption. Therefore, there is an urgent need for a prediction model that can effectively capture complex dynamics while possessing efficient computational capabilities. Against this backdrop, the long-term series model DLinear has attracted increasing attention due to its concise and efficient structure. DLinear significantly reduces computational complexity and improves training and inference speed by decomposing time series into trend and seasonal components and using simple linear layers for feature extraction. Furthermore, DLinear performs well on multiple benchmark tasks, demonstrating its great potential in time series prediction. However, due to the inherent limitations of its linear modeling, DLinear still has shortcomings when facing complex nonlinear trends and high-order seasonal variations. Summary of the Invention

[0005] This invention provides a short-term load forecasting method and system based on trend and seasonal modeling. By using a linear layer with adaptive weight adjustment to extract trend features from load data and by using an extended long short-term memory network to extract seasonal features from load data, it achieves the capture of complex nonlinear trends and high-order seasonal changes, enabling ultra-high accuracy load forecasting within a short timescale.

[0006] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0007] In a first aspect, the present invention provides a short-term load forecasting method based on trend and seasonal modeling, comprising:

[0008] Acquire the load data to be tested;

[0009] The load data to be measured is preprocessed to obtain standard load data, wherein the standard load data includes load information and its corresponding time information;

[0010] The standard load data is input into a pre-trained short-term load forecasting model to obtain short-term load forecast values;

[0011] The method for obtaining the pre-trained short-term load prediction model includes:

[0012] Obtain the historical load dataset and preprocess each historical load data point to obtain the historical sample set;

[0013] The historical sample set is input into a pre-built short-term load forecasting model to obtain a trained short-term load forecasting model. The pre-built short-term load forecasting model extracts trend features from the load data based on a weighted adaptive adjustment linear layer and extracts seasonal features from the load data based on an extended long short-term memory network.

[0014] Optionally, the pre-built short-term load forecasting model includes a time series decomposition module, a feature extraction module, and a fusion forecasting module connected in sequence.

[0015] The time series decomposition module includes a single-layer exponential moving average model;

[0016] The feature extraction module includes an independent trend prediction module and a seasonal prediction module, which are respectively connected to the exponential moving average model.

[0017] The trend prediction module includes an alternating first linear layer (Linear) and a JacobiKAN-Linear linear layer with adaptively adjusted weights based on the Jacobi matrix KAN method.

[0018] The seasonal prediction module includes an extended long short-term memory network (xLSTM) and a second linear layer connected in sequence. The xLSTM includes a first matrix long short-term memory network (mLSTM), a scalar long short-term memory network (sLSTM), and a second matrix long short-term memory network (mLSTM) with residual connections.

[0019] The fusion prediction module includes a stitching layer, a residual connection, a layer normalization Add&Norm layer, and an output layer connected in sequence.

[0020] Optionally, the preprocessing of the load data to be measured to obtain standard load data includes:

[0021] The measured load data is normalized to obtain the processed standard load data, and the calculation formula is as follows:

[0022] ,

[0023] in, These are standard load data. These are the raw load data. and These are the maximum and minimum values ​​of the original load data, while max and min are the target range after scaling.

[0024] Optionally, inputting the standard load data into a pre-trained short-term load forecasting model to obtain short-term load forecast values ​​includes:

[0025] The standard load data is decomposed into trend data and seasonal data by using the time series decomposition module.

[0026] The trend prediction module and the seasonal prediction module respectively make predictions based on the trend data and the seasonal data to obtain the trend prediction value and the seasonal prediction value.

[0027] The short-term load forecast is obtained by fusing trend forecasts and seasonal forecasts using the fusion forecast module.

[0028] Optionally, the step of performing time-series decomposition on the standard load data through the time-series decomposition module to obtain trend data and seasonal data includes:

[0029] Determine if the length of the input time series is sufficient to completely cover one period. If not, pad the time series as follows:

[0030] ,

[0031] in, It is the filled time series. Represents the last few points in the time series. It is the last point in the time series;

[0032] The complete time series is decomposed using the exponential moving average model to obtain the trend and seasonal components of the load data. The calculation formula is as follows:

[0033] ,

[0034] ,

[0035] ,

[0036] in, It is the exponential moving average at time t. α is the observed value at time t, and α is the smoothing factor, with a value range of (0, 1). It is the exponential moving average at time t-1, X Input Indicates input data, X Trend X Season X represents the trend term and the seasonal term, respectively. Res This represents the residual sequence, and Padding(·) represents filling the sequence elements.

[0037] Optionally, the trend prediction module predicts the trend data to obtain a trend prediction value, including:

[0038] The input trend data is linearly transformed using the first linear layer (Linear), and the calculation formula is as follows:

[0039] ,

[0040] in, It is the output of the first linear layer. , ,..., These are input features. , ,..., is the weight corresponding to the input feature, and b is the bias term;

[0041] The JacobiKAN-Linear method, based on the Jacobi matrix, dynamically adjusts the weight distribution of the trend features after linear transformation using an adaptively adjusted linear layer to obtain the trend prediction value. The calculation formula is as follows:

[0042] ,

[0043] in, It is based on the Jacobi matrix-based KAN method, which adaptively adjusts the weights of the linear layer JacobiKAN-Linear. These are the weighting coefficients, and α and β are parameters used to adjust the shape of the Jacobi polynomial. It is a non-linear mapping of the network. These are network parameters.

[0044] Optionally, the seasonal data is predicted using a seasonal prediction module to obtain seasonal prediction values, including:

[0045] By extending the long short-term memory network xLSTM to extract features from the input seasonal data, a seasonal feature vector containing hidden information is obtained.

[0046] The seasonal feature vector is linearly transformed by the second linear layer (Linear) to obtain the seasonal predicted value.

[0047] Optionally, the step of fusing trend forecasts and seasonal forecasts through a fusion forecasting module to obtain short-term load forecasts includes:

[0048] The trend forecast and seasonal forecast values ​​are stitched together using a stitching layer;

[0049] The spliced ​​forecast values ​​are processed with residual connection and layer normalization (Add&Norm) to obtain the final load forecast values.

[0050] The final load forecast value is output through the output layer.

[0051] Optionally, it also includes: calculating the mean absolute error loss function (MAE) and the mean variance error loss function (MSE) between the short-term load forecast values ​​and the corresponding target labels, and evaluating the performance of the trained short-term load forecast model, wherein:

[0052] The formula for calculating the Mean Absolute Error Loss (MAE) is shown below:

[0053] ,

[0054] The formula for calculating the Mean Variance Error Loss (MSE) is shown below:

[0055] ,

[0056] Where N represents the number of samples in the load dataset, This represents the actual load value of the i-th sample. This represents the predicted load value for the i-th sample.

[0057] Secondly, the present invention provides a short-term load forecasting system based on trend and seasonal modeling, comprising:

[0058] Acquire the load data to be tested;

[0059] The load data to be measured is preprocessed to obtain standard load data, wherein the standard load data includes load information and its corresponding time information;

[0060] The standard load data is input into a pre-trained short-term load forecasting model to obtain short-term load forecast values;

[0061] The method for obtaining the pre-trained short-term load prediction model includes:

[0062] Obtain the historical load dataset and preprocess each historical load data point to obtain the historical sample set;

[0063] The historical sample set is input into a pre-built short-term load forecasting model to obtain a trained short-term load forecasting model. The pre-built short-term load forecasting model extracts trend features based on a linear layer with adaptive weight adjustment and extracts seasonal features based on an extended long short-term memory network.

[0064] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: By decomposing load data to obtain trend and seasonal information, a two-stream neural network is used to input the trend and seasonal information into a pre-built trend prediction module and a seasonal prediction module. The JacobiKAN-Linear linear layer, based on the Jacobi matrix, is used to extract the trend features from the load data, and an extended long short-term memory network (xLSTM) is used to extract the seasonal features from the load data. This enables the capture of complex nonlinear trends and high-order seasonal changes, thereby outputting short-term load forecast values ​​and achieving high-precision load forecasting within a short timescale. Attached Figure Description

[0065] Figure 1 The diagram shown is a flowchart of the short-term load forecasting method based on trend and seasonal modeling in Embodiment 1 of the present invention.

[0066] Figure 2 The diagram shown is a flowchart of the short-term load forecasting method based on trend and seasonal modeling in Embodiment 2 of the present invention.

[0067] Figure 3 The diagram shown is a schematic diagram of using exponential moving average to decompose a time series in one embodiment of the present invention.

[0068] Figure 4 The figure shown is a schematic diagram of the structure of a short-term load forecasting model in one embodiment of the present invention;

[0069] Figure 5 The diagram shown is a similarity comparison result between short-term load forecast values ​​and corresponding target data values ​​in one embodiment of the present invention. Detailed Implementation

[0070] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0071] Example 1

[0072] like Figure 1 As shown, this embodiment provides a short-term load forecasting method based on trend and seasonal modeling, including:

[0073] Acquire the load data to be tested;

[0074] The load data to be measured is preprocessed to obtain standard load data, wherein the standard load data includes load information and its corresponding time information;

[0075] The standard load data is input into a pre-trained short-term load forecasting model to obtain short-term load forecast values;

[0076] The method for obtaining the pre-trained short-term load prediction model includes:

[0077] Obtain the historical load dataset and preprocess each historical load data point to obtain the historical sample set;

[0078] The historical sample set is input into a pre-built short-term load forecasting model to obtain a trained short-term load forecasting model. The pre-built short-term load forecasting model extracts trend features from the load data based on a weighted adaptive adjustment linear layer and extracts seasonal features from the load data based on an extended long short-term memory network.

[0079] Example 2

[0080] Based on Example 1, this example also incorporates the following design.

[0081] like Figure 2 As shown, the method in this embodiment specifically includes the following steps:

[0082] Step 1: Obtain load data and perform MinMaxScaler normalization on the load data to obtain normalized load data. The normalized load data includes time information and load information. The calculation expression for MinMaxScaler normalization of the load data is as follows:

[0083] ,

[0084] in, These are standard load data. These are the raw load data. and These are the maximum and minimum values ​​of the original load data. max and min are the target range after scaling, and the default is [0, 1].

[0085] Step 2: Based on time series decomposition, using the principle of exponential moving average, the processed load data is decomposed into a trend term containing data trend information and a seasonal term containing data seasonal fluctuation information.

[0086] If the length of the sequence cannot completely cover a period, the time series length is insufficient for decomposition. The sequence is padded before and after to meet the decomposition requirements. In this embodiment, the selected dataset can cover the complete period, so data padding is not required.

[0087] See Figure 3 Based on the exponential moving average, load data that meets the time-series decomposition requirements is decomposed into trend and seasonal components. The expressions for the trend and seasonal components in the load data are calculated according to the exponential moving average as follows:

[0088] ,

[0089] ,

[0090] ,

[0091] in, It is the exponential moving average at time t. α is the observed value at time t, and α is the smoothing factor, which takes values ​​in the range (0, 1). It is the exponential moving average at time t-1, X Input Indicates input data, XTrend X Season X represents the trend term and the seasonal term, respectively. Res This represents the residual sequence, and Padding(·) represents filling the sequence elements.

[0092] Step 3: Input the trend and seasonal data into the pre-built modular trend and seasonality prediction model respectively, and make predictions accordingly.

[0093] See Figure 4 The time series is decomposed into a trend term and a seasonal term, which are modeled separately using a two-stream neural network structure. The trend term is modeled through a dynamic linear layer, which combines a first linear layer (Linear) with a JacobiKAN-Linear linear layer with adaptively adjusted weights based on the Jacobi matrix KAN method.

[0094] The expression for the ordinary linear layer is as follows:

[0095] ,

[0096] in, It is the output of a normal linear layer. , ,..., These are input features. , ,..., is the weight corresponding to the feature, and b is the bias term.

[0097] The JacobiKAN-Linear linear layer based on the Jacobi matrix KAN method is a linear layer that combines the Jacobi polynomial and the KAN method. JacobiKAN is a kernel method that uses the Jacobi polynomial as the basis function, which can dynamically adjust the weight distribution, enhance the ability to capture trend changes, and has the ability to adjust parameters to adapt to different data distributions. The mathematical expression of the Jacobi polynomial is as follows:

[0098] ,

[0099] Here, α and β are two parameters that change the distribution of the Jacobi multinomial weights. The value range is [-1, 1].

[0100] The mathematical expression for the JacobiKAN-Linear linear layer, which uses the Jacobi matrix-based KAN method for adaptive weight adjustment, is as follows:

[0101] ,

[0102] in, It is the nth Jacobi polynomial. These are weighting coefficients. It is a non-linear mapping of the network. These are network parameters.

[0103] The seasonal term is processed by combining an Extended Long Short-Term Memory (xLSTM) network with a second linear layer (Linear). The xLSTM network models the nonlinear random fluctuations in the residual sequence, extracting periodic and short-term dynamic features from the time series. The xLSTM network stacks three residual blocks—a first mLSTM, a second mLSTM, and a third mLSTM—to form a complete xLSTM architecture, processing the seasonal term information, and then extracting deep features through a regular linear layer (Linear). The scalar long short-term memory (sLSTM) network adds a scalar update mechanism to the traditional LSTM network, introducing exponential gating and a new storage hybrid technique. Its calculation expression is as follows:

[0104] cell stall

[0105] normalizer state

[0106] hidden stall ,

[0107] cell input ,

[0108] input gate ,

[0109] Forget Gate OR ,

[0110] output gate ,

[0111] in, Indicates the cell state of the sLSTM. This represents the output of the forget gate of the sLSTM. This represents the input gate output of the sLSTM. This indicates that the sLSTM has passed through the activation function's input. This represents the input to the sLSTM at the current time step. This represents the normalized state of the sLSTM. This represents the hidden state of the sLSTM. This indicates the output gate of the sLSTM. This represents the Sigmoid activation function. This represents the tanh activation function. This represents the exponential activation function.

[0112] The matrix long short-term memory (mLSTM) network extends the vector operations of traditional LSTM to matrix operations, using the covariance matrix to update memory units. The calculation expression is as follows:

[0113] cell stall

[0114] normalizer state

[0115] hidden stall ,

[0116] query input

[0117] key input

[0118] value input

[0119] input gate ,

[0120] Forget Gate OR ,

[0121] output gate ,

[0122] in, This indicates the cell state of the mLSTM. This represents the output of the forget gate of the mLSTM. This represents the input gate output of an mLSTM. This indicates the input of the mLSTM after the activation function. This represents the input to the mLSTM at the current time step. This represents the normalized state of mLSTM. This represents the hidden state of mLSTM. This indicates the output gate of the mLSTM. Represents the query vector. Represents the key vector. Represents a value vector.

[0123] The Sigmoid function maps real-valued inputs to the (0,1) interval and is used for binary classification. For a very large negative input, its output value is close to 0; for a very large positive input, its output value is close to 1. Its mathematical expression is as follows:

[0124] ,

[0125] in, This represents the input to the Sigmoid function.

[0126] The Tanh activation function is very similar to the Sigmoid activation function, essentially a larger version of the sigmoid function. Its range is (-1, 1), and its mathematical expression is as follows:

[0127] ,

[0128] in, This represents the input to the Tanh activation function.

[0129] Step 4: Merge the two forecasts to output short-term load forecast values;

[0130] In this study, normalized load data is used as target data, and the similarity between short-term load forecasts and their corresponding target data values ​​is calculated using the mean variance error loss function (MSE) and the mean absolute error loss function (MAE).

[0131] Specifically, the calculation expressions for the Mean Absolute Error Loss Function (MAE) and the Mean Variance Error Loss Function (MSE) are as follows:

[0132] ,

[0133] ,

[0134] Where N is the number of data points in the load dataset. This is the actual load data for the i-th load. This is the i-th load forecast data.

[0135] It should be noted that this embodiment selects electricity load data from a district in Nanjing City from July 1, 2016 to July 1, 2019, spanning three years. The data collection frequency is once per hour, totaling 26,304 time samples. The electricity load data from July 1, 2016 to August 31, 2018 is used as the training set, the electricity load data from September 1, 2018 to December 31, 2018 is used as the validation set, and the electricity load data from January 1, 2019 to July 1, 2019 is used as the test set.

[0136] This embodiment is in Figure 5 The data shows a comparison between the predicted and actual electricity load data for June 18-19, 2019. The mean absolute error loss was 0.0359, and the mean variance error was 0.0024. Overall, the error between the short-term load forecast and the actual value is small.

[0137] This application compares the proposed short-term load prediction method with the prediction methods of Long Short-Term Memory Network-Convolutional Neural Network LSTM-CNN, Reformer, and Long Time Series Model DLinear. The results are shown in Table 1.

[0138] Table 1

[0139]

[0140] As shown in Table 1, the short-term load forecasting method of this application outperforms the LSTM-CNN model, Reformer model, and DLinear model in terms of mean absolute error (MAE) and mean variance error (MSE) on the power load dataset, which is in line with the expected results of this application.

[0141] Example 3

[0142] This embodiment provides a short-term load forecasting system based on trend and seasonal modeling, including:

[0143] Acquire the load data to be tested;

[0144] The load data to be measured is preprocessed to obtain standard load data, wherein the standard load data includes load information and its corresponding time information;

[0145] The standard load data is input into a pre-trained short-term load forecasting model to obtain short-term load forecast values;

[0146] The method for obtaining the pre-trained short-term load prediction model includes:

[0147] Obtain the historical load dataset and preprocess each historical load data point to obtain the historical sample set;

[0148] The historical sample set is input into a pre-built short-term load forecasting model to obtain a trained short-term load forecasting model. The pre-built short-term load forecasting model extracts trend features based on a linear layer with adaptive weight adjustment and extracts seasonal features based on an extended long short-term memory network.

[0149] Example 4

[0150] This embodiment provides a computer storage medium storing a computer program. When the computer program is executed by a processor, it implements the short-term load forecasting method based on trend and seasonal modeling as described in any step of Embodiment 2.

[0151] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0152] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0153] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0154] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0155] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A short term load forecasting method based on trend and seasonal modeling, characterized in that, The method comprises the following steps: obtaining to-be-tested load data; preprocessing the to-be-tested load data to obtain standard load data, wherein the standard load data comprises load information and corresponding time information; inputting the standard load data into a pre-trained short-term load prediction model to obtain a short-term load prediction value; wherein the method for obtaining the pre-trained short-term load prediction model comprises: obtaining a historical load data set and preprocessing each historical load data in the historical load data set to obtain a historical sample set; inputting the historical sample set into a pre-constructed short-term load prediction model to obtain a trained short-term load prediction model, wherein the pre-constructed short-term load prediction model extracts trend features in load data based on a weight self-adaptive adjustment linear layer and extracts seasonal features in load data based on an extended long short-term memory network; the pre-constructed short-term load prediction model comprises a time series decomposition module, a feature extraction module and a fusion prediction module connected in sequence; the time series decomposition module comprises an exponential moving average model; the feature extraction module comprises a trend prediction module and a seasonal prediction module which are independent of each other, and the trend prediction module and the seasonal prediction module are connected to the exponential moving average model; the trend prediction module comprises a first linear layer Linear and a weight self-adaptive adjustment linear layer JacobiKAN-Linear based on a Jacobi matrix which are connected alternately; the seasonal prediction module comprises an extended long short-term memory network xLSTM and a second linear layer Linear connected in sequence, wherein the xLSTM comprises a residual connection of a first matrix long short-term memory network mLSTM, a scalar long short-term memory network sLSTM and a second matrix long short-term memory network mLSTM; the fusion prediction module comprises a concatenation layer, a residual connection and a layer normalization Add&Norm layer and an output layer connected in sequence.

2. The short term load forecasting method based on trend and seasonal modeling as claimed in claim 1, wherein, the preprocessing of the to-be-tested load data to obtain the standard load data comprises: normalizing the to-be-tested load data to obtain processed standard load data, and the calculation formula is: ， wherein is the standard load data, is the original load data, and are the maximum and minimum values of the original load data, max and min are the target range after scaling.

3. The short term load forecasting method based on trend and seasonal modeling as claimed in claim 1, wherein, the inputting of the standard load data into the pre-trained short-term load prediction model to obtain the short-term load prediction value comprises: performing time series decomposition on the standard load data through the time series decomposition module to obtain trend item data and seasonal item data; performing prediction according to the trend item data and the seasonal item data through the trend prediction module and the seasonal prediction module respectively to obtain a trend prediction value and a seasonal prediction value; performing fusion on the trend prediction value and the seasonal prediction value through the fusion prediction module to obtain a short-term load prediction value.

4. The short term load forecasting method based on trend and seasonal modeling as claimed in claim 3, wherein, the time series decomposition on the standard load data through the time series decomposition module to obtain the trend item data and the seasonal item data comprises: determining whether the length of the input time series can completely cover a period, and if not, padding the time series, and the expression is as follows: ， wherein, is the time series after padding, represents the last few points of the time series, is the last point of the time series; decomposing the time series with a complete period through the exponential moving average model to obtain trend item data and seasonal item data in the load data, and the calculation formula is as follows: ， ， ， where, is the exponential moving average at time t, is the observation at time t, and a is the smoothing factor, which is in the range (0, 1), is the exponential moving average at time t-1, and X Input denotes the input data, X Trend , X Season denote the trend and seasonality terms, respectively, X Res denotes the residual series, and Padding(·) denotes padding of the sequence elements.

5. The short term load forecasting method based on trend and seasonal modeling as claimed in claim 3, wherein, The trend item data is predicted by a trend prediction module to obtain a trend prediction value, including: The trend item data is linearly transformed by a first linear layer Linear, and the calculation formula is as follows: ， wherein, is the output of the first linear layer Linear, , ..., is an input feature, , ..., is a weight corresponding to the input feature, and b is a bias term; The weight distribution of the linearly transformed trend feature is dynamically adjusted by a Jacobi matrix-based KAN method weight adaptive adjustment linear layer JacobiKAN-Linear to obtain the trend prediction value, and the calculation formula is as follows: ， wherein, is the output of the weight-adapted linear layer JacobiKAN-Linear, are weight coefficients, and a and b are parameters for adjusting the shape of the Jacobi polynomials, is a non-linear mapping of the network, is a network parameter.

6. The short term load forecasting method based on trend and seasonal modeling as claimed in claim 3, wherein, The seasonal item data is predicted by a seasonal prediction module to obtain a seasonal prediction value, including: The seasonal item data is input into an extended long short-term memory network xLSTM for feature extraction to obtain a seasonal feature vector containing hidden information; The seasonal feature vector is linearly transformed by a second linear layer Linear to obtain the seasonal prediction value.

7. The short term load forecasting method based on trend and seasonal modeling as claimed in claim 3, wherein, The trend prediction value and the seasonal prediction value are fused by a fusion prediction module to obtain a short-term load prediction value, including: The trend prediction value and the seasonal prediction value are spliced by a splicing layer; The spliced prediction value is subjected to residual connection and layer normalization Add&Norm processing to obtain a final load prediction value; The final load prediction value is output by an output layer.

8. The short term load forecasting method based on trend and seasonal modeling as claimed in claim 1, wherein, Further comprising: The mean absolute error loss function MAE and the mean square error loss function MSE between the short-term load prediction value and the corresponding target label are calculated to evaluate the performance of the trained short-term load prediction model, wherein: The calculation formula of the mean absolute error loss function MAE is as follows: ， The calculation formula of the mean square error loss function MSE is as follows: ， wherein N represents the number of samples in the load dataset, represents the actual load value of the i-th sample, represents the predicted load value of the i-th sample.

9. A short term load forecasting system based on trend and seasonal modeling, characterized by, Including: Obtaining the to-be-tested load data; The to-be-tested load data is preprocessed to obtain standard load data, wherein the standard load data includes load information and corresponding time information; The standard load data is input into a pre-trained short-term load prediction model to obtain a short-term load prediction value; The pre-trained short-term load prediction model is obtained by: Obtaining a historical load data set and preprocessing each historical load data in the historical load data set to obtain a historical sample set; The historical sample set is input into a pre-constructed short-term load prediction model to obtain a trained short-term load prediction model, wherein the pre-constructed short-term load prediction model extracts trend features based on a weight adaptive adjustment linear layer and extracts seasonal features based on an extended long short-term memory network; The pre-constructed short-term load prediction model includes a time series decomposition module, a feature extraction module, and a fusion prediction module connected in sequence; The time series decomposition module includes an exponential moving average model; The feature extraction module includes a trend prediction module and a seasonal prediction module independent of each other, and the trend prediction module and the seasonal prediction module are connected to the exponential moving average model; The trend prediction module includes a first linear layer Linear and a Jacobi matrix-based KAN method weight adaptive adjustment linear layer JacobiKAN-Linear connected alternately; The seasonal prediction module comprises an extended long short-term memory network xLSTM and a second linear layer Linear connected in sequence, wherein the xLSTM comprises a residual connection first matrix long short-term memory network mLSTM, a scalar long short-term memory network sLSTM and a second matrix long short-term memory network mLSTM; The fusion prediction module comprises a concatenation layer, a residual connection and layer normalization Add&Norm layer and an output layer connected in sequence.

Images