A long-term power system load forecasting method and system based on multi-scale decomposition fusion

Abstract

This invention relates to the field of power load forecasting technology, and in particular to a long-term power system load forecasting method and system based on multi-scale decomposition and fusion. The method includes: data preprocessing based on time series data; obtaining a multi-scale load feature vector set by multi-scale decomposition and feature embedding of the preprocessed data; performing gated adaptive filtering and attention dual-path fusion based on the multi-scale load feature vector set; performing independent prediction and prediction fusion on the fusion results based on a spatiotemporal attention gating mechanism; and evaluating the fused prediction results. The spatiotemporal attention fusion mechanism for multi-scale prediction results proposed in this invention can adaptively integrate prediction information at different scales, avoiding biases caused by a single scale, further improving the comprehensive performance of long-term forecasting, and is applicable to various power system planning and operation scenarios.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and in particular to a long-term power system load forecasting method and system based on multi-scale decomposition and fusion. Background Technology

[0002] The core challenge of long-term load forecasting lies in how to effectively decompose and synergistically utilize information embedded in data at different scales—that is, how to extract sensitive short-term fluctuation characteristics from fine-grained data, while grasping robust long-term evolutionary trends from coarse-grained data, and accurately modeling the nonlinear interactions between its cyclical components, trend components, and external influencing factors.

[0003] Existing long-term power system load forecasting methods fail to fully consider the multi-scale periodic information contained in power load data. Furthermore, traditional forecasting models based on single-scale or linear assumptions fail to fully capture the complex patterns in load data, ranging from micro-level short-term fluctuations (such as intraday variations) to macro-level long-term trends (such as the impact of holidays) and multiple periodicities (daily, weekly, monthly, quarterly, and annual cycles), thus limiting forecast accuracy and generalization ability. Summary of the Invention

[0004] To address the limitations of existing long-term power system load forecasting methods in handling the multi-scale periodicity, complex temporal correlations, and noise interference of load data, this invention provides a Long-Term Power System Load Forecasting Based on Multi-Scale Decomposition-Fusion (LTSLF-MSDF) method. By constructing a multi-scale decomposition and feature embedding module for load sequences and a multi-scale load feature-gated adaptive filtering-attention dual-path fusion module, the load sequences are decomposed into multi-scale information. The fusion of this multi-scale information is achieved by combining gated adaptive filtering technology and an attention mechanism. This effectively filters out noise interference and fully extracts the multi-scale periodicity information contained in the power load data, thereby improving the model's prediction accuracy.

[0005] In a first aspect, the present invention provides a long-term power system load forecasting method based on multi-scale decomposition and fusion, which adopts the following technical solution:

[0006] A long-term power system load forecasting method based on multi-scale decomposition and fusion includes:

[0007] Time series data is obtained based on the acquired power data;

[0008] Data preprocessing based on time series data;

[0009] By performing multi-scale decomposition and feature embedding on the preprocessed data, a set of multi-scale load feature vectors is obtained.

[0010] Gated adaptive filtering and attention dual-path fusion based on multi-scale load feature vector set;

[0011] Independent prediction and predictive fusion of fusion results are performed based on a spatiotemporal attention gating mechanism;

[0012] The results of the prediction fusion are evaluated.

[0013] Secondly, a long-term power system load forecasting system based on multi-scale decomposition and fusion includes:

[0014] The data acquisition module is configured to obtain time-series data based on the acquired power data;

[0015] The preprocessing module is configured to perform data preprocessing based on time series data;

[0016] The feature vector module is configured to obtain a set of multi-scale load feature vectors by performing multi-scale decomposition and feature embedding on the preprocessed data.

[0017] The multi-scale fusion module is configured to perform gated adaptive filtering and attention dual-path fusion based on the multi-scale load feature vector set;

[0018] The prediction fusion module is configured to perform independent prediction and prediction fusion of the fusion results based on a spatiotemporal attention gating mechanism.

[0019] The evaluation module is configured to evaluate the results of the predicted fusion.

[0020] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the long-term power system load forecasting method based on multi-scale decomposition and fusion.

[0021] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide the aforementioned method for long-term power system load forecasting based on multi-scale decomposition and fusion.

[0022] In summary, the present invention has the following beneficial technical effects:

[0023] (1) By performing multi-scale decomposition on the load sequence, this invention effectively separates the periodic patterns and trend components at different time granularities, overcomes the shortcomings of traditional single-scale prediction models in capturing complex time-series features, and significantly improves the accuracy of long-term load prediction.

[0024] (2) The present invention adopts a strategy that combines gated adaptive filtering with attention mechanism to achieve cross-scale fusion of trend features and periodic features. It maintains the consistency of macro trends and makes full use of the detailed information of fine-scale data with reduced noise, thereby enhancing the model’s ability to represent complex electricity consumption behavior and its predictive robustness.

[0025] (3) The spatiotemporal attention fusion mechanism of multi-scale prediction results proposed in this invention can adaptively integrate prediction information at different scales, avoid deviations caused by single scale, further improve the comprehensive performance of long-term prediction, and is applicable to various power system planning and operation scenarios. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a long-term power system load forecasting method based on multi-scale decomposition and fusion according to Embodiment 1 of the present invention;

[0027] Figure 2 This is a schematic diagram of the multi-scale load feature-gated adaptive filtering trend feature fusion operation in Embodiment 1 of the present invention;

[0028] Figure 3 This is a schematic diagram of the multi-scale load feature-attention cycle feature fusion operation in Embodiment 1 of the present invention;

[0029] Figure 4 This is a line graph of the MAE prediction results of the ECL dataset in Embodiment 1 of the present invention;

[0030] Figure 5 This is a line graph of the MSE prediction results for the ECL dataset in Embodiment 1 of the present invention;

[0031] Figure 6 This is a line graph of the RMSE prediction results of the ECL dataset in Embodiment 1 of the present invention;

[0032] Figure 7 This is a line graph of the MAPE prediction results for the ECL dataset in Embodiment 1 of the present invention;

[0033] Figure 8 This is a line graph of the MAE prediction results for the GEFcom2012 dataset in Embodiment 1 of the present invention;

[0034] Figure 9 This is a line graph of the MSE prediction results for the GEFcom2012 dataset in Embodiment 1 of the present invention;

[0035] Figure 10 This is a line graph of the RMSE prediction results for the GEFcom2012 dataset in Embodiment 1 of the present invention;

[0036] Figure 11 This is a line graph of the MAPE prediction results for the GEFcom2012 dataset in Embodiment 1 of the present invention. Detailed Implementation

[0037] The present invention will be further described in detail below with reference to the accompanying drawings.

[0038] Example 1

[0039] Reference Figure 1 This embodiment of a long-term power system load forecasting method based on multi-scale decomposition and fusion includes:

[0040] Preparation of S1 dataset

[0041] S1.1 Electricity Load Forecasting Dataset Acquisition and Processing Flow:

[0042] The Electricity dataset used in this study was obtained through a standardized process. The original data originated from anonymized customer electricity consumption data provided by the Portuguese national grid operator, and was subsequently compiled and published in the UCI Machine Learning Repository by academic institutions. A baseline version was obtained through the standard interface of the UCI Machine Learning Repository; this dataset is widely recognized in the field of energy forecasting research. The data acquisition process followed standardized protocols, accessing the official data directory of the UCI Machine Learning Repository and locating the permanent identifier link for the Electricity Load Diagrams dataset. The download process employed HTTPS security to ensure data integrity and reliable transmission. After obtaining the compressed data package, an integrity check was performed to verify that the file size and checksum matched the official records.

[0043] S1.1.1 Format Conversion and Standardization Processing:

[0044] The original data is in semicolon-separated text format, using the Portuguese standard decimal representation (commas as decimal separators). The conversion process includes the following key steps: reading the original text file, recognizing semicolons as field separators, and converting the comma-separated decimal format to the international standard decimal format. Using commas as field separators conforms to RFC 4180. All numeric fields are converted to floating-point numbers, preserving original precision. The timestamp column is set as the first column in the CSV file for easy time-series analysis. A header row is added to clarify the meaning of each column's data.

[0045] S1.1.2 Final Output Specification:

[0046] Filename: electricity_load_data.csv

[0047] Column structure: Timestamp + 320 customer electricity consumption characteristics + Total electricity consumption target (OT)

[0048] S1.2 GEFCom 2012 Load Forecasting Dataset Acquisition and Processing Flow

[0049] The GEFCom 2012 dataset originates from the global energy forecasting competition hosted by the IEEE Power and Energy Society and is a recognized benchmark dataset in the field of energy forecasting. The competition data package was obtained through the official IEEE data distribution platform. The package is stored in ZIP compressed format and contains multiple CSV files and related metadata. The "load_history.csv" file was extracted as the key data source, recording historical electricity load data for each region.

[0050] S1.2.1 Format Conversion and Standardization Processing:

[0051] The conversion process includes the following steps: Multi-file integration and structure unification: Identifying and integrating scattered regional load data files, standardizing the timestamp format, ensuring cross-regional data alignment, and converting regional identifiers into explicit column headers. CSV format optimization: Constructing a wide table format, with each row containing 20 regional load values ​​for a single time point. The timestamp column serves as the primary key to ensure uniqueness. Numerical fields are standardized to kilowatt units, and unit symbols are removed. Metadata integration: Adding detailed column descriptions to the first row of the CSV file, retaining the original data's time range and regional identification information, and adding comment rows indicating the data source and conversion date.

[0052] S1.2.2 Final Output Specification:

[0053] Filename: gefcom2012_load_forecasting.csv

[0054] Column structure: timestamp + 20 regional load value columns

[0055] S2 dataset preprocessing supplementary preprocessing computation process

[0056] This invention utilizes time-series data from the power sector, encompassing target variables (e.g., load) and multidimensional covariates (e.g., region), and employs instance normalization to eliminate dimensional differences and preserve internal patterns within the series. This processing can be formulated as follows:

[0057] ,

[0058] ,

[0059] in, The target variable at time t is represented. Represents the standardized target variable. Let i represent the i-th covariate at time t. This represents the original eigenvalues. The result obtained after transformation and These represent the mean and standard deviation within the instance, respectively. and and represent the mean and standard deviation of all values ​​of the i-th feature in the current sequence instance, respectively.

[0060] To support the learning of complex long-term dependencies, the dataset integrates three types of key information: historical target sequences, multivariate covariates, and complex time features extracted through encoding, thereby providing reliable data support for power load forecasting.

[0061] Multiscale decomposition and feature embedding of S3 load sequences

[0062] Electricity load sequences at different scales naturally exhibit differentiated characteristics: fine-scale sequences primarily present detailed features, while macro-scale sequences highlight overall changes. This multi-scale perspective can effectively analyze the complex changes of multiple components, thereby improving the modeling effect of electricity load sequences. In forecasting tasks, due to the different dominant patterns at different time scales, multi-scale time series demonstrate differentiated predictive capabilities. Therefore, this invention proposes a multi-scale decomposition and feature embedding method for load sequences to capture the micro-fluctuations and macro-trends of load data and effectively separate the complex mixed features, requiring processing of the original load sequence. To this end, this method employs Multi-scale Decomposition and Feature Embedding (MDFE) operations, which involve decomposing the load sequence into a multi-scale decomposition and feature embedding sequence with a length of... load sequence ( To generate a set of multi-scale load feature vectors, two operations are performed sequentially: Multi-Scale Load Decomposition (MSLD) and Multi-Scale Load Sequence Feature Embedding (Embedding). The operation process of MDFE can be written as follows:

[0063] ,

[0064] in, Represents a set of multi-scale load sequences. This represents the total number of levels in the sequence. This module enables subsequent models to simultaneously consider both the micro-fluctuations and macro-trends of the load data, while also separating features that are difficult to distinguish in the original load data. This facilitates the understanding of complex electricity consumption behaviors contained in the load data by subsequent prediction models. The following sections will detail the specific processes of Multiscale Load Sequence Decomposition (MSLD) and Multiscale Load Sequence Feature Embedding.

[0065] S3.1 Multiscale decomposition of load sequences

[0066] To extract electricity consumption patterns at different time scales from an electricity load sequence, a multi-scale decomposition (MSLD) operation must first be performed on the load sequence. This operation aims to decompose the original sequence into sub-components of different scales (such as periodic and trend components) through average pooling, thereby decoupling micro-fluctuations and macro-trends and improving the robustness and prediction accuracy of time series modeling. To ensure that the multi-scale decomposition can accurately capture and separate the real-world periodic fluctuation patterns in the electricity load that are synchronized with human activities and natural laws, a set of sampling step sizes is set according to the natural period alignment principle:

[0067] ,

[0068] in, This represents five time scales, from fine to coarse: daily, weekly, monthly, quarterly, and yearly. Given an electricity load sequence. ( For time step, (Assuming it's a feature dimension), it is decomposed into a set of multi-scale load sequences through average pooling:

[0069] ,

[0070] in, This represents a set of multi-scale load sequences, implemented using multi-step average pooling operations. , where x is the original input sequence. Sequences at various scales. The formula for average pooling is:

[0071] ,

[0072] in, Indicated by Step-size power load sequence Perform average pooling. The specific calculation process is as follows:

[0073] ,

[0074] in, The length of the downsampled sequence.

[0075] The multi-scale sequence set X obtained after decomposition will serve as the key data basis for subsequent analysis, supporting more accurate time series pattern recognition and load forecasting, enabling the model to take into account both short-term fluctuations and long-term trends.

[0076] S3.2 Multiscale Load Sequence Feature Embedding

[0077] After successfully obtaining the multi-scale load sequence set X, to address the issue of inconsistent feature dimensions between sequences of different scales, we introduce a multi-scale load sequence feature embedding operation. This operation maps load sequences at each scale to a unified high-dimensional feature representation, providing standardized and information-enhanced input features for subsequent deep learning models. The specific computation process is detailed below:

[0078] ,

[0079] in, This represents the set of multi-scale load feature vectors obtained after the embedding operation. Represents a specific scale The corresponding feature vector ( (This represents the time step corresponding to the feature vector). This represents the dimension parameter of the hidden layer in the model. Feature embedding is specifically implemented through linear projection, with the following expansion:

[0080] ,

[0081] in, It is the learnable weight matrix at the m-th scale. It is the corresponding bias term. Indicates the dimension of the input features. This represents the dimension parameter of the hidden layer in the model. This indicates the process through a linear transformation layer: The input sequence at each scale Projecting from its original feature space to a high-dimensional model hidden space. Wherein, This represents the transpose of the learnable weight matrix at the m-th scale.

[0082] This embedding operation allows us to effectively separate and extract features from the raw load data that are difficult to distinguish and identify directly. This greatly facilitates the subsequent prediction model's in-depth understanding and accurate grasp of the complex electricity consumption behavior contained in the load data. This step plays a crucial role in the overall data processing flow, laying a solid foundation for the training and prediction of subsequent models.

[0083] S4 Multiscale Load Feature Gated Adaptive Filtering - Attention Dual-Path Fusion

[0084] The Multi-scale Load Feature Extraction with a Gated Adaptive Filtering-Attention Dual-Path Fusion (MLFEGAFADPF) operation involves sequentially performing load feature vector decomposition, multi-scale load feature-gated adaptive filtering trend feature fusion, multi-scale load feature-attention periodic feature fusion, and multi-scale data fusion to generate a reconstructed multi-scale load feature vector set. Historical observation data presents complex real-world time series, exhibiting both clear periodicity and trends. Notably, periodicity and trend have different characteristics in time series analysis, corresponding to short-term and long-term variations, or stationary and non-stationary dynamic processes, respectively. In each layer of the load feature vector decomposition module, load decomposition is performed on the input multi-scale time series, separating periodic and trend features. Subsequently, the decomposed periodic and trend features are independently fused at multiple scales. The trend feature fusion employs adaptive filtering combined with a top-down strategy, where coarse-scale trends guide the formation of fine-scale trends, while adaptive filtering removes noise from the fine-scale data. This approach ensures that the model consistently follows the overall evolution direction of the sequence when predicting short-term changes, without allowing noise at the fine scale to affect the accuracy of the prediction results. The fusion of periodic features employs an attention mechanism combined with a bottom-up strategy. Specifically, it dynamically and selectively weights and fuses fine-scale periodic information using an attention mechanism to form coarse-scale periodic information. Simultaneously, this module enhances the analysis of complex time series relationships, improving the robustness and accuracy of the prediction results.

[0085] Based on the above process, the module performs a fusion operation on historical information at different scales. For the l-th layer, its input is... The cross-scale fusion process of MLFEGAFADPF (Multi-scale Load Feature Gated Adaptive Filtering-Attention Dual-Path Fusion) can be denoted as:

[0086] ,

[0087] Where L represents the depth of historical information extraction by the model.

[0088] Indicates length is Feature dimension is This module reconstructs a set of multi-scale load feature vectors. Through refined decomposition and fusion of multi-scale load features, it can not only effectively capture the periodic changes in load data at different scales, but also accurately identify its trend patterns. This processing method greatly improves the comprehensiveness and accuracy of load data feature extraction, directly enhancing the subsequent prediction model's ability to learn complex load patterns and its prediction accuracy.

[0089] S4.1 Load Eigenvector Decomposition

[0090] Load Feature Vector Decomposition (LFVD) employs a "divide and conquer" strategy. Specifically, it involves decomposing the multi-scale load feature vectors of the (l-1)th layer into a single, multi-scale model. Decomposed into a set of periodic load eigenvectors and load trend eigenvector set The periodic eigenvectors contain repetitive fluctuations in the load data related to daily, weekly, and monthly cycles, while the trend eigenvectors reflect the slowly changing long-term trends in the load data. The load eigenvector decomposition operation of the l-th layer can be denoted as:

[0091] ,

[0092] in , indicating the first Load eigenvectors at scale m in the layer This indicates a load characteristic decomposition operation. and They represent the first The load periodicity feature vector and load trend feature vector are decomposed at scale m in the layer. This operation extracts the load trend feature vector through Discrete Fourier Transform (DFT):

[0093] ,

[0094] in For the first Historical load observations at the m-scale within the layer, This represents the Discrete Fourier Transform. Based on this, the operation obtains the load periodicity feature vector by subtracting the load trend feature vector at the corresponding scale from the load feature vector:

[0095] ,

[0096] The trend feature vectors and periodic feature vectors generated by the load feature vector decomposition operation are then processed in the scale load feature-filtered trend feature fusion module and the multi-scale load feature-attention periodic feature fusion module, respectively.

[0097] S4.2 Multi-scale load characteristics - gated adaptive filtering trend feature fusion

[0098] The trend term represents the long-term macroscopic evolution of the sequence. At coarse scales (e.g., monthly or yearly), it contains more stable and global information, while at fine scales (e.g., daily), it is susceptible to local fluctuations and noise. Fusion from coarse to fine scales allows the macroscopic trend at the coarse scale to be passed down as "prior knowledge," correcting local biases at the fine scale, thereby achieving trend alignment across multiple scales and enhancing the model's consistent modeling ability. Although the trend component is predominantly low-frequency, it still contains high-frequency noise (e.g., measurement errors or sudden disturbances), and direct fusion would reduce prediction robustness. Gated adaptive filtering dynamically adjusts the filtering strength through a gating mechanism, adaptively preserving the true trend signal while filtering noise: on the one hand, it avoids the over-smoothing problem caused by traditional fixed filtering; on the other hand, the gating mechanism allows the model to differentiate the processing of information in different frequency bands according to scale characteristics (e.g., coarse scales focus on global smoothing, while fine scales preserve local details).

[0099] For trend characteristics The residual connection method is used for the first The trend components at each scale are fused from top to bottom. The fusion of trend features is carried out while keeping the coarse-scale trend features unchanged. Then, the coarse-scale trend features are processed and added to the fine-scale trend features. Trend information is passed down through the long-term trend, ensuring the consistency of the prediction direction. In the layer The cross-scale fusion method for load trend characteristics at different scales can be denoted as:

[0100] ,

[0101] ,

[0102] in, This represents a coarse-to-fine scale fusion technique incorporating gated adaptive filtering, with an input dimension of... The output dimension is The fusion process utilizes coarse-scale trend characteristics Dimensionality reduction is performed to reduce the sequence length from... Compressed to a length equal to the target fine scale m Consistency is achieved, thus aligning features across scales. The aligned coarse-scale features are then fed into a gated adaptive filtering unit, which processes them through a dual-path parallel process: the filtering path generates basic trend components through linear transformation, while the gated path generates dynamic weight coefficients using the Sigmoid function. The two paths are multiplied element-wise to form the gated adaptive filter output. and This represents the learnable weight matrix, used to perform a linear transformation on the fused features. and This indicates that the learnable bias vector is used in conjunction with the weight matrix. This represents element-wise multiplication. The generated characteristics... It will be passed as input to the data fusion module for further fusion processing.

[0103] S4.3 Multiscale Load Feature-Attention Cycle Feature Fusion

[0104] In the periodic feature fusion stage, an attention-guided fine-scale to coarse-scale fusion strategy is adopted to effectively integrate high-frequency detail information contained in the fine-scale load sequence into the coarse-scale representation, thereby enhancing the integrity of the coarse-scale periodic modeling. Specifically, this method first preserves the periodic features at the finest scale, then uses an attention mechanism to weight the fine-scale periodic features and fuses them into adjacent coarse-scale features. The attention mechanism adaptively evaluates the importance of different fine-scale features to coarse-scale generation, thus enabling the selection and enhancement of key details. This fine-to-coarse fusion direction helps retain richer local periodic fluctuation information in the coarse-scale representation, allowing it to reflect subtle changes while maintaining the macroscopic periodic structure. Technically, for periodic features… The residual connection method is used to perform fine-scale to coarse-scale fusion of periodic components to achieve fine-scale to coarse-scale interaction of multi-scale periodic information. The cross-scale fusion method of load periodic characteristics at scale m in layer l can be denoted as:

[0105] ,

[0106] ,

[0107] ,

[0108] This indicates that a fusion strategy from fine-scale to coarse-scale is used, with an input dimension of... The output dimension is First, regarding fine-scale periodic features... By applying an attention mechanism, a weighted output is calculated for the query (Q), key (K), and value (V), where, This represents the calculation of the dot product of the query matrix and the transpose of the key matrix, the result of which is an attention score matrix representing the relevance of each query to all keys. Represents the actual information of all elements. First, the dimensions of the query matrix and key matrix are considered. Second, an upsampling operation is used to reduce the weighted feature dimensions from a finer scale. Align to coarse scale To ensure sequence length compatibility, the processed features are then integrated into coarse-scale periodic features using residual connections. The generated features... It will be passed as input to the data fusion module for further fusion processing.

[0109] S4.4 Multiscale Data Fusion

[0110] When dealing with complex and variable load data, data feature extraction and fusion are key steps in building efficient data analysis models. We first extract a series of multi-scale periodic feature sets from the load data, which can be represented as... , where m ranges from 0 to M-1, covering various scales from the shortest to the longest period. Similarly, to comprehensively capture the dynamic trends of the data, a multi-scale trend feature set was extracted, which is represented as . It reflects the changing trends of data at different time scales.

[0111] After extracting the periodic and trend feature sets, multi-scale data fusion becomes the core of the entire analysis process. This step aims to achieve deep fusion of features of different scales and properties to construct a more information-rich and structurally unified feature representation. The fusion module not only effectively integrates periodic and trend features but also adaptively adjusts the information aggregation strategy according to different time periods and contexts, thus ensuring that the final feature representation possesses both global trend consistency and local detail resolution. The multi-scale data fusion of the l-th module can be formulated as follows:

[0112] ,

[0113] It is a feedforward neural network module, consisting of two linear layers and a GELU activation function. Its input is periodic features. and trend characteristics The result after initial fusion is the enhanced feature representation. Then, Output and original input By combining features through residual connections to preserve low-level feature information and avoid gradient vanishing, an enhanced feature representation that integrates multi-scale periodic and trend information is generated after the fusion operation. This feature not only comprehensively captures the complex patterns and structures in the data, but also possesses strong generalization ability and predictive performance. It can provide strong support for subsequent data analysis, model building, and decision support.

[0114] S5 Multi-Scale Prediction Results: Spatiotemporal Attention Fusion

[0115] The core of this module lies in the introduction of a spatiotemporal attention gating mechanism. This mechanism adaptively captures complex spatiotemporal dependencies in sequences at different scales, achieving accurate fusion of independent prediction results at multiple scales. Spatiotemporal attention gating dynamically assesses the importance of prediction results at each scale in the spatiotemporal dimension and performs weighted integration accordingly, thereby effectively mining the multi-periodic characteristics inherent in the load sequence. This mechanism not only significantly improves the model's prediction accuracy and generalization ability but also enhances the model's adaptability and robustness to abrupt loads and complex time-series patterns.

[0116] S5.1 Multi-scale independent load forecasting

[0117] Because power load sequences exhibit significant pattern heterogeneity across different time scales, a segmented specialized modeling approach is adopted, configuring a dedicated predictor for each scale to achieve targeted pattern learning and feature extraction, thereby ensuring prediction accuracy at each scale from the modeling source. By integrating the advantages of multi-scale prediction through a fusion module, the robustness and prediction accuracy of the overall model can be systematically improved. This operation combines the multi-scale load feature vectors fused from the previous layer... As input. To accommodate the unique characteristics of sequences at different scales, the feature vectors for each scale are input into a dedicated predictor to generate independent prediction results for the corresponding scale:

[0118] ,

[0119] in, Indicates batch size as The time step is The number of features is The Scale prediction results, This represents a dedicated predictor for the m-th scale, and the formula for calculating the predictor is: ,in, This is the output of the multi-scale data fusion module. w and b are the weight matrix and bias vector of the linear layer, respectively. After obtaining the independent prediction results at each scale, a spatiotemporal attention weight matrix is ​​first constructed based on the spatiotemporal attention mechanism to adaptively evaluate the contribution of the prediction results at each scale to the final prediction. Subsequently, the independent prediction results at each scale are weighted and fused according to the calculated attention weights.

[0120] These specialized predictors can be optimized for the characteristics of load data at different scales, thereby capturing unique load variation patterns at each scale. For example, for daily-scale data, the predictor may focus more on intraday load fluctuations; while for weekly or monthly-scale data, it may focus more on capturing periodic trends. In this way, multi-scale independent load forecasting operations can generate more accurate and detailed forecast results, providing a solid foundation for subsequent spatiotemporal attention-gated fusion.

[0121] S5.2 Multi-scale Prediction Results: Spatiotemporal Attention Gating Fusion

[0122] After obtaining independent prediction results at each scale, the spatiotemporal attention gating fusion operation of the multi-scale prediction results calculates the attention weights of the prediction results at each scale in the spatiotemporal dimension. These weights reflect the relative importance of the prediction results at different scales in spatiotemporal dimensions. Specifically, by constructing a spatiotemporal attention mechanism, the model can adaptively evaluate the contribution of the prediction results at each scale to the final prediction.

[0123] Based on the calculated attention weights, the independent prediction results at each scale are weighted and fused. This fusion method not only considers the accuracy of the prediction results at each scale but also fully considers their spatiotemporal correlation. Through spatiotemporal attention-gated fusion, the model can effectively integrate multi-scale information, thereby generating more accurate and comprehensive load prediction results. The calculation of the spatiotemporal attention weight matrix can be denoted as:

[0124] ,

[0125] The calculation of the spatiotemporal attention weight matrix A is a deep learning process involving multi-step nonlinear transformations, which integrates prediction results at various scales. The tensors are concatenated along the feature dimensions to form a tensor integrating multi-scale information. Then, a flattening operation is performed to reshape the tensor into a two-dimensional matrix to meet the processing requirements of the fully connected layer, thereby achieving joint modeling of the spatiotemporal dimensions. The concatenated tensor undergoes two consecutive linear transformation layers: first, by the weight matrix... and bias Feature projection is performed, and nonlinear expressive power is introduced through the ReLU activation function; subsequently, the weight matrix is ​​used... and bias The features are mapped to the scale dimension. Finally, the output is normalized using the softmax function to generate matrix A, where... Indicates batch size as The time step is The number of scales is The spatiotemporal attention weight matrix, This represents a multi-scale prediction tensor concatenated along the feature dimensions. and Represents the learnable spatiotemporal attention weight matrix. and This represents the bias vector. The operation reshapes the input tensor into a two-dimensional matrix to facilitate spatiotemporal joint modeling. Finally, the independent predictions at each scale are weighted and fused, as shown below:

[0126] ,

[0127] in, This indicates the final prediction result after fusion. This represents a slice of the spatiotemporal attention weight tensor A at the m-th scale. This represents element-wise multiplication. Indicates batch size as The time step is The number of features is The Scale prediction results.

[0128] The final prediction result integrates feature information from multi-scale load data across different temporal and spatial dimensions, and achieves adaptive weighted fusion of prediction results at each scale through a spatiotemporal attention gating mechanism. This fusion method not only fully considers the multi-period characteristics of the load sequence, but also effectively improves the model's ability to capture complex time-series patterns by dynamically adjusting the contribution weights of prediction results at each scale. In practical applications, this prediction result can provide more accurate and reliable load forecasting basis for power system scheduling and planning, helping to optimize resource allocation, reduce operating costs, and improve the overall stability and robustness of the system.

[0129] S6 Power Load Prediction Experiment

[0130] To verify the effectiveness of the LTSLF-MSDF method in long-term prediction tasks, this study conducted a series of comprehensive experiments. This method demonstrates significant advantages over current state-of-the-art models in long-term prediction tasks, as evidenced by its outstanding performance on multiple benchmark tests.

[0131] S6.1 Experimental Dataset Selection,

[0132] This study conducts multivariate long-term load forecasting based on two power datasets: ECL (Electricity) and GEFCom2012. We performed instance normalization on the data and constructed the dataset by dividing it into a 7:1:2 time series partition. The core task requires the model to integrate historical target sequences, multivariate covariates, and complex temporal features to accurately capture long-term dependencies, ultimately achieving accurate forecasts for the next 96 to 720 time steps. Detailed dataset descriptions are shown in Table 1.

[0133] Table 1: Detailed Data of Dataset Samples

[0134]

[0135] ECL (Electricity): This dataset records the hourly electricity consumption (kW) of 320 customers in Portugal. It fully covers the special period of daylight saving time and exhibits significant periodic fluctuations, making it particularly suitable for electricity load forecasting, energy efficiency analysis, and demand-side management research. It provides a high-value real-world benchmark for model validation.

[0136] GEFCom2012: This dataset primarily comprises two relatively independent tracks: load forecasting and wind power forecasting. We selected the load_history dataset for load forecasting as the forecast data, which records hourly load history data (in kW) for 20 regions. This dataset, with its high-quality, real-world multi-regional load data, has become an authoritative benchmark for evaluating model performance in the field of power forecasting.

[0137] S6.2 Evaluation Criteria

[0138] This study uses four evaluation metrics to measure predictive performance: mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and mean absolute percentage error (MAPE). Specifically, MAE measures the average magnitude of the prediction error; MSE, using squaring, is more sensitive to larger errors; RMSE, as the square root of MSE, maintains the same unit of measurement as the original data, facilitating result interpretation; and MAPE, presented as a percentage, is suitable for evaluating relative accuracy.

[0139] MAE: ,

[0140] MSE: ,

[0141] RMSE: ,

[0142] MAPE: ,

[0143] T: Represents the total number of time steps or the total number of data points. : Represents the actual value at time point t. : Represents the predicted value at time point t.

[0144] S6.3 Selection of Comparison Method

[0145] To comprehensively verify the superior performance of the LTSLF-MSDF method, this experiment selected representative mainstream methods in the field of long-term time series forecasting as control objects. These include Transformer-type models that have performed exceptionally well in time series forecasting tasks in recent years, such as Informer, Autoformer, Crossformer, iTransformer, and FEDformer; it also compared them with some cutting-edge improved variant models, such as TiDE and PatchTST. These control methods cover different network architecture design philosophies, demonstrating the unique advantages of the LTSLF-MSDF model in capturing load sequence patterns from multiple dimensions.

[0146] S6.4 Hyperparameter Settings

[0147] All experiments were implemented using the PyTorch framework and run on an NVIDIA RTX 4090 graphics card. For long-term prediction tasks, we used average pooling to decompose the original sequence into subsequences of different scales, and fixed the input sequence length to a uniform 96. The batch size was 16, and the total number of training epochs was 20.

[0148] To ensure fair comparison, all model hyperparameters were determined through validation set system tuning. The parameter combination ultimately adopted by the model in this paper is the configuration that achieved the best performance on the validation set under this optimization process. These values ​​collectively shape a relatively lightweight and efficient model architecture that is sufficient to capture the key features of the data while effectively avoiding overfitting, thus achieving excellent generalization performance on prediction tasks of various step sizes (96 to 720 steps), specifically reflected in low MAE (mean absolute error), MSE (mean squared error), RMSE (root mean squared error), and MAPE (mean absolute percentage error) indices. For the model proposed in this paper, its core hyperparameter settings are as follows: hidden layer dimension is 16, number of encoding layers is 3, number of decoding layers is 1, number of attention heads is set to 8, the optimizer uses an initial learning rate of 0.001, the weight decay coefficient is 1e-5, and an early stopping mechanism is introduced—a patience value of 15 is set, and if the validation set loss (MSE) does not decrease for several consecutive rounds, training is terminated early. Predictions were made for the next 96, 192, 336, and 720 steps, and the corresponding MAE (mean absolute error), MSE (mean squared error), RMSE (root mean squared error), and MAPE (mean absolute percentage error) values ​​were obtained. The smaller the four values, the higher the prediction performance. Detailed hyperparameter settings are shown in Table 2.

[0149] Table 2: Hyperparameter settings of the comparison models

[0150] Model Hyperparameter settings LTSLF-MSDF Encoder layers = 3; Decoder layers = 1; Multi-head attention heads = 8; Batch size = 16; Number of iterations = 20 Autoformer Encoder layers = 2; Decoder layers = 1; Multi-head attention heads = 8; Batch size = 16; Number of iterations = 20 Informer Encoder layers = 2; Decoder layers = 1; Multi-head attention heads = 8; Batch size = 16; Number of iterations = 20 Crossformer Encoder layers = 2; Decoder layers = 1; Multi-head attention heads = 8; Batch size = 16; Number of iterations = 20 TiDE Encoder layers = 2; Decoder layers = 2; Multi-head attention heads = 8; Batch size = 16; Number of iterations = 20 iTransformer Encoder layers = 3; Decoder layers = 1; Multi-head attention heads = 8; Batch size = 16; Number of iterations = 20 PatchTST Encoder layers = 2; Decoder layers = 1; Multi-head attention heads = 8; Batch size = 16; Number of iterations = 20 FEDformer Encoder layers = 2; Decoder layers = 1; Multi-head attention heads = 8; Batch size = 16; Number of iterations = 20

[0151] Please refer to Figure 1 , Figure 1 This is a schematic diagram of the overall architecture of a long-term power system load forecasting method based on multi-scale decomposition and fusion. From... Figure 1 As can be seen from the above, the method proposed in this invention mainly comprises four core components: a multi-scale decomposition and feature embedding module for load sequences, a multi-scale load feature-gated adaptive filtering-attention dual-path fusion module, and a spatiotemporal attention-gated fusion module for multi-scale prediction results. This method decomposes the power load sequence at multiple scales, extracting its periodic and trend features at different time granularities (e.g., daily, weekly, monthly, quarterly, and annually), and achieves cross-scale information fusion using attention mechanisms, adaptive filtering techniques, and feedforward fusion strategies, thereby significantly improving the accuracy and robustness of long-term load forecasting.

[0152] The multi-scale decomposition and feature embedding module of the load sequence decomposes the original load sequence into subsequences of multiple scales by setting multiple time resolutions (such as step size set S= {1, 7, 31, 91, 365}) and using average pooling operation. Then, it maps these subsequences into high-dimensional feature vectors through the embedding layer, thereby achieving effective decoupling and expression of micro fluctuations and macro trends in the load data.

[0153] The multi-scale load feature gated adaptive filtering-attention dual-path fusion module further decomposes the multi-scale load features, extracting periodic and trend features separately, and employs a combined top-down and bottom-up strategy for cross-scale fusion. Specifically, the trend feature fusion uses gated adaptive filtering (GAF) technology. Figure 2 Using coarse-scale trends to guide the modeling of fine-scale trends effectively suppresses noise interference; periodic feature fusion utilizes an attention mechanism to achieve dynamic weighted aggregation of fine-scale periodic information to coarse-scale information. Figure 3 This allows for the enhancement of the ability to characterize complex electricity consumption behaviors while maintaining predictive consistency.

[0154] The multi-scale data fusion method deeply fuses the periodic and trend features obtained after dual-path fusion. This module uses a bilinear feedforward network with a GELU activation function to achieve inter-channel information interaction and feature recombination, further optimizing feature representation and providing a unified and strengthened feature foundation for the final prediction.

[0155] The multi-scale prediction result spatiotemporal attention-gated fusion module performs spatiotemporal attention-weighted fusion of independent prediction results at each scale to generate the final load prediction result. This module introduces a spatiotemporal attention gating mechanism to dynamically evaluate the importance of predictions at each scale in the spatiotemporal dimension and perform weighted integration accordingly, thereby further improving the accuracy, robustness and generalization ability of long-term predictions.

[0156] To verify the effectiveness of the LTSLF-MSDF model in long-term power load forecasting, we conducted extensive experiments. The experimental results (see Tables 3-6) are presented below. Figures 4-11 The LTSLF-MSDF model demonstrates superior performance compared to current state-of-the-art models in long-term prediction tasks, consistently outperforming other models across various benchmark tests covering diverse datasets with varying numbers of variables and real-world application scenarios. This fully validates its effectiveness.

[0157] In summary, the long-term power system load forecasting method and system based on multi-scale decomposition and fusion (LTSLF-MSDF) proposed in this invention can effectively address the nonlinearity, non-stationarity, and multi-scale characteristics of power load sequences, significantly improve forecast accuracy, and provide a reliable basis for power grid planning, energy dispatch, and market decision-making.

[0158] Tables 3-6 show the prediction results of the LTSLF-MSDF model. The model was compared with seven state-of-the-art models on two datasets. The experimental metrics were MAE, MSE, RMSE, and MAPE. The experimental results are as follows.

[0159] Table 3: MAE Prediction Results

[0160]

[0161] Table 4: MSE Prediction Results

[0162]

[0163] Table 5: RMSE Prediction Results

[0164]

[0165] Table 6: MAPE Prediction Results

[0166]

[0167] Based on the combined results in Tables 3 (MAE), 4 (MSE), 5 (RMSE), and 6 (MAPE), the model of this invention (Ours) achieved best or near-best performance on most prediction lengths (96, 192, 336, and 720 steps) and on both datasets (ECL and GEFCom2012). The advantage was particularly pronounced at longer prediction step lengths (such as 336 and 720), indicating that the model has strong long-term dependency capture ability and generalization ability.

[0168] from Figures 4-11 As can be seen, the errors of all models increase with the increase in prediction length, but the increase in the error of the model of this invention (Ours) is the most gradual. This indicates that the model of this invention has better stability when facing long-term predictions. In particular, on the GEFCom2012 dataset, the MAE of the LTSLF-MSDF model at 720 prediction steps is only 0.629, which is much lower than that of Autoformer (0.715) and Informer (1.039), indicating that LTSLF-MSDF can effectively suppress error accumulation.

[0169] Example 2

[0170] This embodiment provides a long-term power system load forecasting system based on multi-scale decomposition and fusion, including:

[0171] The data acquisition module is configured as a computer-readable storage medium storing multiple instructions adapted to be loaded and executed by the processor of a terminal device as described in the long-term power system load forecasting method based on multi-scale decomposition and fusion.

[0172] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement various instructions; the computer-readable storage medium being configured to store multiple instructions adapted for loading and execution by the processor of the aforementioned long-term power system load forecasting method based on multi-scale decomposition and fusion.

[0173] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A long-term power system load forecasting method based on multi-scale decomposition fusion, characterized in that, The method comprises the following steps: obtaining time series data based on the acquired power data; data preprocessing based on the time series data; obtaining a multi-scale load feature vector set through multi-scale decomposition and feature embedding of the preprocessed data; gate adaptive filtering and attention dual-path fusion under multi-scale load features based on the multi-scale load feature vector set; independent prediction and prediction fusion of the fusion result based on the spatio-temporal attention gate mechanism; evaluation of the prediction fusion result; The multi-scale load feature vector set is obtained by performing multi-scale decomposition and feature embedding on the preprocessed data, including adopting multi-scale decomposition and feature embedding operations, by dividing the length of the load sequence , the number of features C The multi-scale load feature vector set is generated by sequentially performing multi-scale decomposition MSLD and multi-scale load sequence feature embedding on the load sequence , and setting a sampling step set: wherein, denotes the five-level time scales of day, week, month, season, and year from fine to coarse scales, and the given power load sequence , is the time step, is the feature dimension, which is decomposed into a multi-scale load sequence set by the average pooling operation: wherein, } denotes a multi-scale load sequence set, which is realized by a multi-step average pooling operation, , x is the original input sequence, and each scale The average pooling calculation formula is: wherein, denotes an average pooling operation over the power load sequence with a step size of ; the specific calculation process is as follows: wherein, is the down-sampled sequence length, after obtaining the multi-scale load sequence set X, a multi-scale load sequence feature embedding mechanism is introduced, which is represented as: wherein, denotes a set of multi-scale load feature vectors obtained after the embedding operation, , } denotes a certain scale the corresponding feature vector, denotes the time step corresponding to the feature vector, denotes the dimension parameter of the hidden layer in the model, and the specific implementation of the feature embedding is completed by linear projection, and the expansion formula is: where, is the m-th scale learnable weight matrix, is the corresponding bias term, denotes the dimension of the input feature, denotes the dimension parameter of the hidden layer in the model, denotes a linear transformation layer: projects each scale input sequence from its original feature space to a high-dimensional model hidden space, where, denotes the transpose of the m-th scale learnable weight matrix; The MLFEGAFADPF based on the multi-scale load feature vector set under the multi-scale load feature includes sequentially performing load feature vector decomposition, gate adaptive filtering trend feature fusion, and attention period feature fusion, and multi-scale data fusion on the multi-scale load feature vector set, and finally generating an enhanced load feature vector set that fuses multi-scale period and trend information. For the lth layer, the input is The cross-scale fusion process of the MLFEGAFADPF is recorded as: ， Where L represents the depth of historical information extraction by the model. Indicates length is Feature dimension is The reconstructed multi-scale load feature vector set is then subjected to load feature vector decomposition (LFVD). Based on the load feature vector decomposition, the multi-scale load feature vector set of the (l-1)th layer is... Decomposed into a set of periodic load eigenvectors and load trend eigenvector set The periodic feature vector contains repetitive fluctuations in the load data related to daily, weekly, and monthly cycles, while the trend feature vector reflects the slowly changing long-term trend in the load data. The load feature vector decomposition operation for the l-th layer is as follows: ， in, Indicates the first Load feature vector at the layer scale This indicates a load characteristic decomposition operation. and They represent the first The load periodicity feature vector and load trend feature vector after decomposition at scale m in the layer, where the decomposition operation extracts the load trend feature vector through Discrete Fourier Transform (DFT), are expressed as: ， wherein is the first mesoscale historical load observation value in the layer, denotes the discrete Fourier transform, on the basis of which the load periodicity eigenvector is obtained by subtracting the load trend eigenvector corresponding to the scale from the load eigenvector: ， trend feature vectors and periodic feature vectors generated through load feature vector decomposition are respectively processed through gate adaptive filtering trend feature fusion and attention periodic feature fusion; The gated adaptive filtering trend feature fusion includes adopting adaptive filtering combined with a fusion strategy from coarse scale to fine scale, wherein for trend features , a residual connection mode is adopted to implement top-down fusion of trend components of the first scale, then the coarse scale trend features are processed and added to the fine scale trend features, the long-term trend is passed down to transfer trend information, and the consistency of the prediction direction is ensured, and the cross-scale fusion method of the load trend features in the first scale is represented as: ​ ， ， wherein, denotes the self-coarse-to-fine scale fusion incorporating gated adaptive filtering technique, with input dimension and output dimension , the fusion process utilizes to reduce the dimensionality of the coarse scale trend feature , compressing the sequence length from to the length of the target fine scale m , thus realizing the alignment of cross-scale features, the aligned coarse scale feature is then fed into the gated adaptive filtering unit and processed through a dual-path parallel processing: the filtering path generates the base trend component through linear transformation, while the gating path produces dynamic weight coefficients with the help of Sigmoid function, both are multiplied element-wise to form the gated adaptive filtering output, wherein, and denote the learnable weight matrices for linear transformation of the fused features, and denote the learnable bias vectors used in conjunction with the weight matrices, denotes the element-wise multiplication; The attention cycle feature fusion includes adopting an attention mechanism to combine a self-fine-to-coarse scale fusion method to integrate information from a fine scale time sequence to a coarse scale, wherein, for the cycle feature A residual connection method is adopted to implement self-fine-to-coarse scale fusion of the periodic component of the mth scale: to achieve self-fine-to-coarse scale interaction of multi-scale seasonal information, and a cross-scale fusion method of the load cycle feature of the mth scale in the lth layer is represented as: ， ， ， denotes the fusion strategy from fine scale to coarse scale, the input dimension is , the output dimension is , firstly, the fine scale periodic feature is applied with attention mechanism to calculate the weighted output of query, key and value, where denotes the calculation of dot product of query matrix and transpose of key matrix, represents the actual information of all elements, is the dimension of query matrix and key matrix, secondly, the dimension of weighted feature is aligned from fine scale to coarse scale by upsampling operation to ensure the compatibility of sequence length, finally, the processed feature is integrated into coarse scale periodic feature by means of residual connection.

2. The long-term power system load forecasting method based on multi-scale decomposition fusion according to claim 1, characterized in that, the data preprocessing based on the time series data comprises the following steps: ， ， wherein, denotes the target variable at time t, denotes the normalized target variable, denotes the i-th covariate at time t, denotes the original feature value the result after transformation, denote the mean and standard deviation within the instance, respectively, and denote the mean and standard deviation of all values of the i-th feature in the current sequence instance, respectively.

3. The long-term power system load forecasting method based on multi-scale decomposition fusion according to claim 2, characterized in that, The multi-scale data fusion includes fusing the input periodic features and trend features to construct a unified feature representation, wherein the multi-scale data fusion of the lth module can be formulated as: ， is a feed-forward neural network module consisting of two linear layers and a GELU activation function, whose input is the periodic feature and the trend feature After preliminary fusion, the output is an enhanced feature representation, and then, The output of the original input is combined through a residual connection to retain the underlying feature information and avoid gradient disappearance. After fusion, an enhanced feature representation is generated that fuses multi-scale periodic and trend information .

4. The long-term power system load forecasting method based on multi-scale decomposition fusion according to claim 3, characterized in that, The spatio-temporal attention gate mechanism is used for independent prediction and prediction fusion of the fusion result, including introducing a spatio-temporal attention gate mechanism, adaptively capturing complex spatio-temporal dependencies in different scale sequences, and realizing accurate fusion of multi-scale independent prediction results, wherein the multi-scale load independent prediction includes As input, the feature vectors of each scale are respectively input into a dedicated predictor to generate independent prediction results of the corresponding scale in order to adapt to the unique characteristics of different scale sequences. ， in, Indicates batch size as The time step is The number of features is The Scale prediction results, This represents a dedicated predictor for the m-th scale, and the formula for calculating the predictor is: ,in, The output of the multi-scale data fusion module is w, where w and b are the weight matrix and bias vector of the linear layer, respectively. After obtaining the independent prediction results at each scale, a spatiotemporal attention weight matrix is ​​first constructed based on the spatiotemporal attention mechanism to adaptively evaluate the contribution of each scale's prediction results to the final prediction. Subsequently, the independent prediction results at each scale are weighted and fused according to the calculated attention weights. The spatiotemporal attention weight matrix is ​​expressed as follows: ， The calculation of the spatio-temporal attention weight matrix A is a deep learning process including multiple nonlinear transformations, which combines the prediction results of each scale along the feature dimension to form a tensor integrating multi-scale information, and then reshapes the tensor into a two-dimensional matrix through a flattening operation to adapt to the processing requirements of the fully connected layer, thereby realizing joint modeling of the spatio-temporal dimension; the concatenated tensor is subjected to two consecutive linear transformation layers: first, the feature is projected by the weight matrix and the bias , and the nonlinear expression capability is introduced through the ReLU activation function; then the feature is mapped to the scale dimension by the weight matrix and the bias ; finally, the output is normalized by the softmax function to generate the matrix A, where denotes the spatio-temporal attention weight matrix with a batch size of , a time step of , and a scale number of , denotes the multi-scale prediction tensor concatenated along the feature dimension, and denotes the learnable spatio-temporal attention weight matrix, and denotes the bias vector, The flattening operation is used to reshape the input tensor into a two-dimensional matrix for joint spatio-temporal modeling, and finally the independent prediction results of each scale are weighted and fused, represented as: ， wherein, denotes the final prediction result after fusion, denotes a slice of the spatio-temporal attention weight tensor A at the m-th scale, denotes an element-wise multiplication, denotes the prediction result at the m-th scale with batch size B, time step T, and number of features C.

5. The long-term power system load forecasting method based on multi-scale decomposition fusion according to claim 4, characterized in that, identifying and integrating scattered regional load data files, unifying the timestamp format, then optimizing the CSV format, constructing a wide table format, each row containing 20 regional load values of a time point, then adding detailed column descriptions in the first row of the CSV file, retaining the time interval and regional identification information of the original data, adding data sources and conversion date annotation lines, and finally using a normalization method to eliminate dimensional differences and retain sequence internal patterns, expressed as: MAE: , MSE: , RMSE: , MAPE: , T: represents the total number of time steps or total number of data points, : represents the true value at time point t, : represents the predicted value at time point t.

6. A long-term power system load forecasting system based on multi-scale decomposition fusion, performing a long-term power system load forecasting method based on multi-scale decomposition fusion according to claim 1, characterized in that, the evaluation of the prediction fusion result comprises the following steps: selecting mean absolute error MAE to measure the average amplitude of the prediction error; selecting mean square error MSE to perceive the error through square operation; selecting root mean square error RMSE as the square root of MSE for result interpretation; selecting mean absolute percentage error MAPE to present the error in percentage form for relative accuracy evaluation, respectively expressed as: The method comprises the following steps: a data acquisition module configured to obtain time series data based on the acquired power data; a preprocessing module configured to preprocess data based on the time series data; a feature vector module configured to obtain a multi-scale load feature vector set through multi-scale decomposition and feature embedding of the preprocessed data; a multi-scale fusion module configured to perform gate adaptive filtering and attention dual-path fusion under multi-scale load features based on the multi-scale load feature vector set; a prediction fusion module configured to perform independent prediction and prediction fusion of the fusion result based on the spatio-temporal attention gate mechanism; an evaluation module configured to evaluate the prediction fusion result.

Images