Short-term power load prediction method and system based on graph-enhanced time-frequency dual-branch structure

Abstract

The application discloses a short-term power load prediction method and system based on a graph-enhanced time-frequency double-branch structure and belongs to the technical field of power system load prediction. The application firstly pre-processes a power load sequence, obtains phase-enhanced time sequence characteristics through adaptive time domain inquiry, extracts multi-scale enhanced characteristics through multi-scale convolution, and then inputs the time-frequency double-branch to obtain time domain and frequency domain preliminary prediction results respectively; graph structure modeling is further used to realize variable correlation enhancement and obtain graph-enhanced time-frequency characteristics; finally, adaptive weighted fusion is performed to obtain power load prediction results. The application can accurately model load cycle characteristics, fully excavate variable coupling correlation, significantly improve the precision, robustness and scene adaptation capability of short-term load prediction, and meet the engineering application requirements of new power systems.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, and discloses a short-term power load forecasting method and system based on a graph-enhanced time-frequency dual-branch structure. Background Technology

[0002] Short-term power load forecasting is a core foundation for the safe and stable operation and economic dispatch of power systems. Its forecasting accuracy directly affects key decisions such as generator unit allocation, power flow calculation, and reserve capacity configuration. Power load sequences have complex characteristics, including strong nonlinearity, multi-period coupling, multi-variable correlation, and multi-scale fluctuations.

[0003] Current deep learning methods in short-term load forecasting primarily focus on single-branch modeling in the time domain, such as those based on Transformer and LSTM. These methods can only uncover the temporal dependencies of load sequences, neglecting the inherent frequency-domain periodicity and harmonic component characteristics of the load sequences. They cannot simultaneously model the co-modulation of time-domain trend changes and frequency-domain periodic patterns, resulting in insufficient comprehensiveness in feature extraction. Furthermore, they do not adequately explore the multi-scale characteristics of load sequences. Load sequences contain features at multiple time scales, including daily, weekly, and hourly fluctuations. Existing models often employ single-scale convolution or fixed receptive field encoding structures, which cannot effectively capture fluctuation features at different scales, leading to weak predictive capabilities for sudden load fluctuations. Time-frequency feature fusion often uses fixed-weight concatenation or simple summation, failing to adaptively adjust the fusion weights of the time-frequency branches according to the spectral characteristics of different load sequences. This results in poor generalization ability and difficulty in adapting to the load forecasting needs of different scenarios. Summary of the Invention

[0004] This invention addresses the technical problems in existing short-term load forecasting technologies, such as incomplete mining of time-frequency features, insufficient capture of multi-scale characteristics, inadequate utilization of variable correlations, and lack of adaptability in fusion methods. Therefore, this invention provides a short-term power load forecasting method and system based on a graph-enhanced time-frequency dual-branch structure.

[0005] To achieve the above-mentioned technical effects, the technical solution adopted by this invention is: a short-term power load forecasting method based on a graph-enhanced time-frequency dual-branch structure, comprising the following steps: S1. Preprocess the input power load sequence to obtain a standardized sequence; perform adaptive time-domain query on the standardized sequence to obtain a phase-enhanced time-series feature representation; S2. Perform multi-scale convolutional feature extraction on the temporal feature representation to obtain multi-scale enhanced features; S3. Input the multi-scale enhancement features into the time domain branch and the frequency domain branch respectively for processing to obtain the preliminary prediction results in the time domain and the preliminary prediction results in the frequency domain. S4. Perform graph-based variable correlation enhancement on the preliminary time-domain prediction results and the preliminary frequency-domain prediction results respectively to obtain graph-enhanced time-domain features and graph-enhanced frequency-domain features. S5. Adaptively weightedly fuse the graph-enhanced time-domain features and graph-enhanced frequency-domain features to obtain the final power load prediction result.

[0006] As a preferred embodiment, step S1 includes: Z-score standardization is performed on all numerical features of the power load sequence, and the standardized power load sequence is divided into training set, validation set and test set according to time order to obtain the divided power load sequence. For the partitioned power load sequence, initialize the periodic parameter matrix used for periodic modeling in the adaptive time-domain query, and the periodic parameter matrix is ​​initialized using Xavier uniform initialization; The relative index of each time step within the corresponding period length of the period parameter matrix is ​​calculated by modular arithmetic, and the query vector is retrieved from the period parameter matrix based on the relative index to generate a query tensor; The query tensor is used as the query vector, and the embedding features of the normalized sequence are used as keys and values. The aggregated context information is calculated through a scaling dot product attention mechanism. The aggregated context information and the embedded features of the standardized sequence are fused together through residual connections to obtain the temporal feature representation.

[0007] As a preferred embodiment, the expression for the periodic parameter matrix is: ; The range of uniform initialization of Xavier is: ; The formula for the modular arithmetic operation is as follows: ; The formula for calculating the scaled dot product attention is: ; Among them, Q param For the periodic parameter matrix, L cycle idx is a predefined period length, D is the feature dimension; t For the relative index of the t-th time step within the period, I cycle The starting period index for the standardized sequence; For the aggregated context information matrix, To query a tensor; Embedding features for standardized sequences; for Perform transpose operation on the last two dimensions; This is the scaling factor.

[0008] As a preferred embodiment, step S2 includes: The temporal feature representation is processed by a parallel multi-branch one-dimensional convolutional structure. The temporal feature representation is input into each branch of the one-dimensional convolutional structure to obtain the convolutional output feature of each branch. In the parallel multi-branch one-dimensional convolutional structure, each branch is set with a convolutional kernel of a different size; The convolutional output features are sequentially subjected to batch normalization, ReLU activation, and Dropout processing to obtain the preprocessed features of each branch. The preprocessed features are concatenated along the channel dimension to obtain a multi-scale joint feature map; The multi-scale joint feature map is fused and its dimensionality is compressed by a one-dimensional convolutional layer with a kernel size of 1 to obtain a fused feature with the same channel dimension as the temporal feature representation. The fused features and the temporal feature representations are fused through residual connections to obtain multi-scale enhanced features.

[0009] As a preferred embodiment, step S3 includes: The multi-scale enhanced features are input into a temporal coding network based on the iTransformer architecture for processing to obtain preliminary temporal prediction results; The multi-scale enhanced features are input into the frequency domain modeling network for processing to obtain preliminary frequency domain prediction results.

[0010] As a preferred embodiment, the graph-based variable correlation enhancement in step S4 specifically includes: Average pooling is performed on the feature tensors corresponding to the preliminary prediction results in the time domain and the preliminary prediction results in the frequency domain respectively to obtain the variable-level feature representations for each variable. The variable-level feature representation is subjected to L2 norm normalization, and the cosine similarity between variables is calculated based on the normalized variable-level feature representation to generate an initial similarity matrix. The initial similarity matrix is ​​sparsified to obtain a sparse similarity matrix. An adjacency matrix is ​​constructed by adding an identity matrix to the sparse similarity matrix. Row normalization is then performed on the adjacency matrix using the degree matrix to obtain a normalized adjacency matrix. Based on the normalized adjacency matrix, graph convolution feature aggregation is performed on the feature tensors corresponding to the preliminary prediction results in the time domain and the feature tensors corresponding to the preliminary prediction results in the frequency domain, respectively, to obtain graph convolution time domain features and graph convolution frequency domain features; Nonlinear transformations are performed on the graph convolutional time-domain features and graph convolutional frequency-domain features respectively, and then fused with the corresponding preliminary prediction results through residual connections to obtain the graph-enhanced time-domain features and graph-enhanced frequency-domain features.

[0011] As a preferred embodiment, the adaptive weighted fusion in step S5 specifically includes: Perform a discrete Fourier transform on the standardized sequence to obtain the spectral information corresponding to the sequence. Calculate the energy ratio of the main frequency and harmonic components based on the spectral information to obtain the time-domain fusion weight and frequency-domain fusion weight. Based on the time-domain fusion weights and frequency-domain fusion weights, the graph-enhanced time-domain features and graph-enhanced frequency-domain features are weighted and summed to obtain the final power load prediction result.

[0012] Based on the above method, the present invention also proposes a short-term power load forecasting system based on a graph-enhanced time-frequency dual-branch structure, comprising: The data preprocessing module is used to preprocess the input power load sequence to obtain a standardized sequence; and to perform adaptive time-domain query on the standardized sequence to obtain a phase-enhanced time-series feature representation. A multi-scale feature extraction module is used to perform multi-scale convolutional feature extraction on the temporal feature representation to obtain multi-scale enhanced features; The dual-branch processing module is used to input the multi-scale enhancement features into the time-domain branch and the frequency-domain branch respectively for processing, so as to obtain the preliminary prediction results in the time domain and the preliminary prediction results in the frequency domain. The graph enhancement module is used to enhance the variable correlation of the preliminary time-domain prediction results and the preliminary frequency-domain prediction results based on graph structure, respectively, to obtain graph-enhanced time-domain features and graph-enhanced frequency-domain features. The fusion prediction module is used to adaptively weight and fuse the graph-enhanced time-domain features and graph-enhanced frequency-domain features to obtain the final power load prediction result.

[0013] As a preferred embodiment, the data preprocessing module includes: The standardization unit is used to perform Z-score standardization on all numerical features of the power load sequence, and divide the standardized power load sequence into training set, validation set and test set in chronological order to obtain the divided power load sequence. The periodic parameter initialization unit is used to initialize the periodic parameter matrix used for periodic modeling in the adaptive time-domain query for the divided power load sequence. The periodic parameter matrix is ​​initialized using Xavier uniform initialization. The query tensor generation unit is used to calculate the relative index of each time step within the corresponding period length of the period parameter matrix through modular arithmetic, retrieve the query vector from the period parameter matrix based on the relative index, and generate a query tensor. The context information aggregation unit is used to take the query tensor as a query vector, the embedding features of the normalized sequence as keys and values, and calculate the aggregated context information through a scaling dot product attention mechanism. The temporal feature representation generation unit is used to fuse the aggregated context information with the embedded features of the standardized sequence through residual connections to obtain the temporal feature representation.

[0014] As a preferred embodiment, the graph enhancement module includes: The variable-level feature generation unit is used to perform average pooling processing on the feature tensor corresponding to the preliminary prediction result in the time domain and the feature tensor corresponding to the preliminary prediction result in the frequency domain, respectively, to obtain the variable-level feature representation corresponding to each variable. The similarity calculation unit is used to perform L2 norm normalization on the variable-level feature representation, calculate the cosine similarity between variables based on the normalized variable-level feature representation, and generate an initial similarity matrix. The adjacency matrix construction unit is used to perform sparsification on the initial similarity matrix to obtain a sparse similarity matrix, add an identity matrix to the sparse similarity matrix to construct an adjacency matrix, and perform row normalization on the adjacency matrix through the degree matrix to obtain a normalized adjacency matrix. The graph convolution unit is used to perform graph convolution feature aggregation on the feature tensors corresponding to the preliminary prediction results in the time domain and the feature tensors corresponding to the preliminary prediction results in the frequency domain, based on the normalized adjacency matrix, to obtain graph convolution time domain features and graph convolution frequency domain features. The graph enhancement feature generation unit is used to perform nonlinear transformations on the graph convolutional time-domain features and graph convolutional frequency-domain features respectively, and then fuse them with the corresponding preliminary prediction results through residual connection to obtain the graph enhancement time-domain features and graph enhancement frequency-domain features.

[0015] Compared with existing technologies, this invention achieves accurate modeling and phase enhancement of the periodicity of load sequences through an adaptive time-domain query mechanism based on the periodic parameter matrix and scaled dot product attention. This effectively solves the problem of insufficient capture of load periodic characteristics by traditional methods and improves the expressive power of time-series features.

[0016] By employing a parallel, multi-branch, multi-scale convolutional structure, different-sized convolutional kernels are used to capture the fluctuation characteristics of load sequences at different time scales. Combined with residual connections, feature fusion is achieved, which effectively improves the model's ability to capture sudden load fluctuations and multi-cycle coupling characteristics.

[0017] A dual-branch coding network in the time and frequency domains is constructed. The time-domain branch uses iTransformer to capture long-distance temporal dependencies, while the frequency-domain branch mines the periodicity and harmonic characteristics of the sequence. This achieves collaborative modeling of time-domain trends and frequency-domain patterns, and solves the problem of incomplete feature extraction in single-branch modeling.

[0018] By constructing a graph structure based on cosine similarity and aggregating graph convolutional features, we can deeply mine the coupling correlation between multiple load variables, integrate the correlation information between variables into the feature expression, and further improve the prediction accuracy.

[0019] The time-frequency fusion weights are adaptively calculated based on the spectral characteristics of the load sequence, replacing the traditional fixed-weight fusion method. This method can dynamically adapt to load sequences with different scenarios and characteristics, greatly improving the model's generalization ability and scenario adaptability. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall process logic of the short-term power load forecasting method based on a graph-enhanced time-frequency dual-branch structure of the present invention. Figure 2 This is a schematic diagram of the module structure of the short-term power load forecasting system based on the graph-enhanced time-frequency dual-branch structure of the present invention; Figure 3 This is a schematic diagram of the data preprocessing module structure of the short-term power load forecasting system based on graph-enhanced time-frequency dual-branch structure of the present invention; Figure 4 This is a schematic diagram of the graph enhancement module structure of the short-term power load forecasting system based on the graph enhancement time-frequency dual-branch structure of the present invention; in: 1. Data preprocessing module; 11. Standardization unit; 12. Periodic parameter initialization unit; 13. Query tensor generation unit; 14. Context information aggregation unit; 15. Temporal feature representation generation unit; 2. Multi-scale feature extraction module; 3. Dual-branch processing module; 4. Graph Enhancement Module; 41. Variable-Level Feature Generation Unit; 42. Similarity Calculation Unit; 43. Adjacency Matrix Construction Unit; 44. Graph Convolution Unit; 45. Graph Enhancement Feature Generation Unit; 5. Fusion prediction module. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the embodiments and accompanying drawings. However, this should not be construed as limiting the scope of the above-described subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0022] refer to Figure 1 Example 1: A short-term power load forecasting method based on a graph-enhanced time-frequency dual-branch structure. The specific implementation process is as follows: Step S1: Data Preprocessing and Adaptive Time-Domain Query The input power load sequence is standardized and preprocessed, and the phase enhancement of the time series features is achieved through adaptive time domain query, fully covering all the technical features defined in claims 2 and 3.

[0023] First, the input power load sequence is preprocessed: the power load sequence is a multivariate time series containing power load values ​​and related influencing factors, including but not limited to meteorological characteristics, time characteristics, and power grid operation characteristics; Z-score standardization is performed on all numerical features of the power load sequence to eliminate the influence of different units on model training; the standardized power load sequence is divided into training set, validation set, and test set according to time order, where the training set is used for model parameter learning, the validation set is used for model hyperparameter tuning and early shutdown control, and the test set is used for model performance evaluation; a sliding window method is used to construct samples from the divided sequence to generate sample sequences that meet the model input requirements, the sample sequences including historical input sequences and output sequences to be predicted.

[0024] Secondly, adaptive time-domain querying is performed on the standardized sequence to generate a phase-enhanced temporal feature representation: ① Initialization of the periodic parameter matrix: Based on the inherent periodic characteristics of the power load sequence, the corresponding period length L is predefined. cycle Initialize the periodic parameter matrix Q used for periodic modeling. param The dimension of the periodic parameter matrix is ​​L. cycle ×D, where D is the feature dimension of the input sequence; The periodic parameter matrix is ​​initialized using Xavier uniform initialization, with an initialization range of [value missing]. This ensures the uniform distribution of initialization parameters and avoids gradient vanishing or gradient exploding.

[0025] ② Query tensor generation: The relative index of each time step within the period length is calculated using modular arithmetic. The formula for modular arithmetic is: ; Among them, idx t For the relative index of the t-th time step within the period, I cycle The initial periodic index is used to standardize the sequence; based on the calculated relative index, the query vector at the corresponding position is retrieved from the periodic parameter matrix to generate a query tensor that matches the length of the input sequence.

[0026] ③ Contextual Information Aggregation: The generated query tensor is used as the query vector for the attention mechanism, and the embedding features of the normalized sequence are used as the keys and values ​​for the attention mechanism. The aggregated contextual information is calculated using a scaled dot product attention mechanism. The formula for calculating scaled dot product attention is: ; in, For the aggregated context information matrix, To query a tensor; Embedding features for standardized sequences; for Perform transpose operation on the last two dimensions; This is a scaling factor, and its value is consistent with the feature dimension D; By using a scaled dot product attention mechanism, it is possible to adaptively aggregate contextual information that matches the periodic phase of the current time step, thereby enhancing the periodic phase characteristics of the sequence.

[0027] ④ Residual fusion: The aggregated context information matrix and the embedded features of the standardized sequence are fused through residual connection to retain the basic features of the original sequence, while the phase-enhanced context information is superimposed to obtain the phase-enhanced temporal feature representation.

[0028] Step S2: Multi-scale convolution feature extraction Multi-timescale feature extraction and fusion are performed on temporal feature representations to obtain multi-scale enhanced features.

[0029] ① Multi-branch parallel convolution processing: Construct a parallel multi-branch one-dimensional convolution structure, with each branch having a one-dimensional convolution layer, and the kernel sizes of different branches being different, to capture the sequence fluctuation features at different time scales; input the temporal feature representation obtained in step S1 into each branch of the one-dimensional convolution structure, and obtain the convolution output features of each branch through convolution operation.

[0030] ② Convolutional Feature Preprocessing: For the convolutional output features of each branch, batch normalization, ReLU activation, and Dropout are performed sequentially. Batch normalization is used to accelerate model convergence and reduce internal covariate bias. ReLU activation is used to introduce nonlinear transformations to improve the model's feature fitting ability. Dropout is used to randomly deactivate some neurons to avoid overfitting. Finally, the preprocessed features of each branch are obtained.

[0031] ③ Multi-scale feature stitching: The preprocessed features of all branches are stitched together along the channel dimension to fuse feature information from different time scales and obtain a multi-scale joint feature map.

[0032] ④ Feature fusion and dimensionality compression: Through a one-dimensional convolutional layer with a kernel size of 1, cross-channel feature fusion and dimensionality compression are performed on the multi-scale joint feature map, so that the channel dimension of the output feature is consistent with the channel dimension of the input temporal feature representation, and the fused feature is obtained.

[0033] ⑤ Residual fusion: The fused features are fused with the original temporal feature representation through residual connections to avoid the gradient vanishing problem caused by deep networks, while retaining the basic temporal features and multi-scale enhanced features, and finally obtaining multi-scale enhanced features.

[0034] Step S3: Time-frequency dual-branch parallel processing By performing parallel modeling of multi-scale enhancement features through time-domain and frequency-domain branches respectively, preliminary prediction results in the time domain and frequency domain are obtained.

[0035] ① Temporal branching: Multi-scale enhanced features are input into a temporal encoding network based on the iTransformer architecture. The temporal encoding network contains at least one iTransformer encoder layer, and each encoder layer contains a multi-head self-attention module and a feedforward neural network module. By mapping the variable dimension to the query, key, and value space of the attention mechanism through the iTransformer architecture, the long-range temporal dependencies of multivariate time series and the global correlations between variables can be captured more accurately. Finally, the feature tensor corresponding to the preliminary temporal prediction results is output.

[0036] ② Frequency Domain Branching: The multi-scale enhanced features are input into the frequency domain modeling network. The frequency domain modeling network first performs a discrete Fourier transform on the input features to convert the time-domain features into frequency-domain features, separating different frequency components in the sequence. Then, the frequency-domain attention module adaptively weights the frequency-domain features to highlight the characteristic information of the main frequency and harmonic components related to the load cycle characteristics and suppress noise components. Finally, the inverse discrete Fourier transform converts the weighted frequency-domain features into time-domain features, and finally outputs the feature tensor corresponding to the preliminary frequency domain prediction result.

[0037] Step S4: Enhance variable correlation based on graph structure Graph structure modeling is performed on the preliminary prediction results in the time domain and frequency domain respectively to explore the coupling correlation between variables and achieve feature enhancement.

[0038] For both the feature tensors corresponding to the preliminary prediction results in the time domain and the feature tensors corresponding to the preliminary prediction results in the frequency domain, the same processing procedure is used to perform variable correlation enhancement based on graph structure. The specific process is as follows: ① Variable-level feature generation: Perform global average pooling on the input feature tensor in the time dimension to eliminate the influence of the time dimension and obtain a one-dimensional variable-level feature representation for each variable.

[0039] ② Calculation of similarity between variables: Perform L2 norm normalization on the variable-level feature representation to eliminate the influence of feature magnitude on similarity calculation; based on the normalized variable-level feature representation, calculate the cosine similarity between each pair of variables to generate an initial similarity matrix. The element values ​​of the initial similarity matrix represent the degree of association between the corresponding two variables.

[0040] ③ Adjacency Matrix Construction and Normalization: The initial similarity matrix is ​​sparsified, retaining only the high similarity associations corresponding to each variable and setting the remaining low similarity associations to zero, resulting in a sparse similarity matrix. This avoids noise interference and computational redundancy caused by fully connected graphs. An identity matrix is ​​added to the sparse similarity matrix to retain the feature information of the variables themselves, constructing a graph-structured adjacency matrix. The degree matrix corresponding to the adjacency matrix is ​​calculated, and row normalization is performed on the adjacency matrix using the degree matrix to obtain a normalized adjacency matrix, ensuring the numerical stability of graph convolution operations.

[0041] ④ Graph Convolution Feature Aggregation: Based on the normalized adjacency matrix, graph convolution feature aggregation is performed on the input feature tensor. The feature information of adjacent nodes is aggregated through graph convolution operation to realize information interaction between variables and obtain graph convolution features.

[0042] ⑤ Residual fusion and feature enhancement: The graph convolutional features are subjected to nonlinear transformation through a fully connected layer to further mine the deep information of the aggregated features; then the nonlinearly transformed features and the feature tensors corresponding to the preliminary prediction results of the original input are fused through residual connections to finally obtain the graph-enhanced features, i.e., the time domain branch corresponds to the graph-enhanced time domain features, and the frequency domain branch corresponds to the graph-enhanced frequency domain features.

[0043] Step S5: Adaptive Weighted Fusion and Final Prediction Adaptive weighted fusion of graph-enhanced time-domain features and graph-enhanced frequency-domain features yields the final power load prediction result.

[0044] ① Adaptive fusion weight calculation: Perform a discrete Fourier transform on the standardized sequence obtained in step S1 to obtain the spectral information corresponding to the sequence; calculate the proportion of the total energy of the main frequency and all harmonic components in the sequence to the total energy of the sequence based on the spectral information. This proportion is the frequency domain fusion weight. The time domain fusion weight is the difference between 1 and the frequency domain fusion weight, and the sum of the time domain fusion weight and the frequency domain fusion weight is 1. In this way, the fusion weights of the time domain and frequency domain features can be dynamically adjusted according to the spectral characteristics of the input sequence itself to achieve adaptive fusion.

[0045] ② Weighted fusion and final prediction: Based on the calculated time-domain fusion weights and frequency-domain fusion weights, the graph-enhanced time-domain features and graph-enhanced frequency-domain features are weighted and summed to obtain the fused final features; the final features are then mapped to an output sequence that matches the prediction length through a fully connected layer, thus obtaining the final power load prediction result.

[0046] refer to Figure 2 , Figure 3 Example 2: A short-term power load forecasting system based on a graph-enhanced time-frequency dual-branch structure is used to implement the above forecasting method. The specific implementation methods of each module of the system are as follows: Data preprocessing module 1: Performs all operations of step S1 above, preprocesses the input power load sequence to obtain a standardized sequence, and performs adaptive time-domain query on the standardized sequence to obtain a phase-enhanced time-series feature representation. The data preprocessing module 1 includes a standardization unit 11, a period parameter initialization unit 12, a query tensor generation unit 13, a context information aggregation unit 14, and a time-series feature representation generation unit 15, all connected in sequence. Standardization Unit 11: Used to perform Z-score standardization on all numerical features of the power load sequence, and divide the standardized power load sequence into training set, validation set and test set in time order to obtain the divided power load sequence; Periodic parameter initialization unit 12: used to initialize the periodic parameter matrix for periodic modeling in adaptive time domain query on the divided power load sequence, wherein the periodic parameter matrix is ​​initialized uniformly by Xavier; Query tensor generation unit 13: used to calculate the relative index of each time step within the corresponding period length of the period parameter matrix through modular arithmetic, and to retrieve the query vector from the period parameter matrix based on the relative index to generate the query tensor; Context information aggregation unit 14: used to take the query tensor as the query vector, the embedding features of the normalized sequence as the key and value, and calculate the aggregated context information through the scaling dot product attention mechanism; Temporal feature representation generation unit 15: used to fuse the aggregated context information with the embedded features of the standardized sequence through residual connections to obtain a temporal feature representation.

[0047] Multi-scale feature extraction module 2: This module performs all operations of step S2 above, extracts multi-scale convolutional features from the temporal feature representation, and obtains multi-scale enhanced features.

[0048] Dual-branch processing module 3: is used to perform all the operations of step S3 above, inputting the multi-scale enhancement features into the time domain branch and the frequency domain branch respectively for processing, and obtaining the preliminary prediction results in the time domain and the preliminary prediction results in the frequency domain.

[0049] Graph enhancement module 4: Performs all operations of step S4 above, enhancing variable correlations based on graph structures on the preliminary time-domain and frequency-domain prediction results to obtain graph-enhanced time-domain features and graph-enhanced frequency-domain features. The graph enhancement module 4 includes a variable-level feature generation unit 41, a similarity calculation unit 42, an adjacency matrix construction unit 43, a graph convolution unit 44, and a graph-enhanced feature generation unit 45, which are sequentially connected in communication. Variable-level feature generation unit 41: is used to perform average pooling processing on the time dimension for the feature tensors corresponding to the preliminary prediction results in the time domain and the feature tensors corresponding to the preliminary prediction results in the frequency domain, respectively, to obtain the variable-level feature representation for each variable; Similarity calculation unit 42: is used to perform L2 norm normalization on the variable-level feature representation, calculate the cosine similarity between variables based on the normalized variable-level feature representation, and generate an initial similarity matrix; Adjacency matrix construction unit 43: is used to perform sparsification on the initial similarity matrix to obtain a sparse similarity matrix, add an identity matrix to construct an adjacency matrix based on the sparse similarity matrix, and perform row normalization on the adjacency matrix through the degree matrix to obtain a normalized adjacency matrix; Graph convolution unit 44: Based on the normalized adjacency matrix, it performs graph convolution feature aggregation on the feature tensors corresponding to the preliminary prediction results in the time domain and the feature tensors corresponding to the preliminary prediction results in the frequency domain, respectively, to obtain graph convolution time domain features and graph convolution frequency domain features. Graph Enhancement Feature Generation Unit 45: This unit performs nonlinear transformations on graph convolutional time-domain features and graph convolutional frequency-domain features respectively, and then fuses them with the corresponding preliminary prediction results through residual connections to obtain graph enhancement time-domain features and graph enhancement frequency-domain features.

[0050] Fusion prediction module 5: It is used to perform all the operations of step S5 above, and to adaptively weight and fuse the graph-enhanced time domain features and graph-enhanced frequency domain features to obtain the final power load prediction result.

[0051] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A short-term power load forecasting method based on a graph-enhanced time-frequency dual-branch structure, characterized in that, Includes the following steps: S1. Preprocess the input power load sequence to obtain a standardized sequence; perform adaptive time-domain query on the standardized sequence to obtain a phase-enhanced time-series feature representation; S2. Perform multi-scale convolutional feature extraction on the temporal feature representation to obtain multi-scale enhanced features; S3. Input the multi-scale enhancement features into the time domain branch and the frequency domain branch respectively for processing to obtain the preliminary prediction results in the time domain and the preliminary prediction results in the frequency domain. S4. Perform graph-based variable correlation enhancement on the preliminary time-domain prediction results and the preliminary frequency-domain prediction results respectively to obtain graph-enhanced time-domain features and graph-enhanced frequency-domain features. S5. Adaptively weightedly fuse the graph-enhanced time-domain features and graph-enhanced frequency-domain features to obtain the final power load prediction result. 2.The short-term power load forecasting method based on graph-enhanced time-frequency dual-branch structure according to claim 1, wherein, Step S1 includes: Z-score standardization is performed on all numerical features of the power load sequence, and the standardized power load sequence is divided into training set, validation set and test set according to time order to obtain the divided power load sequence. For the partitioned power load sequence, initialize the periodic parameter matrix used for periodic modeling in the adaptive time-domain query, and the periodic parameter matrix is ​​initialized using Xavier uniform initialization; The relative index of each time step within the corresponding period length of the period parameter matrix is ​​calculated by modular arithmetic, and the query vector is retrieved from the period parameter matrix based on the relative index to generate a query tensor; The query tensor is used as the query vector, and the embedding features of the normalized sequence are used as keys and values. The aggregated context information is calculated through a scaling dot product attention mechanism. The aggregated context information and the embedded features of the standardized sequence are fused together through residual connections to obtain the temporal feature representation. 3.The short-term power load forecasting method based on graph enhanced time-frequency dual-branch structure according to claim 2, characterized in that, The expression for the periodic parameter matrix is: ； The range of uniform initialization of Xavier is: ； The formula for the modular arithmetic operation is as follows: ； The formula for calculating the scaled dot product attention is: ； Among them, Q param For the periodic parameter matrix, L cycle idx is a predefined period length, D is the feature dimension; t For the relative index of the t-th time step within the period, I cycle The starting period index for the standardized sequence; For the aggregated context information matrix, To query a tensor; Embedding features for standardized sequences; for Perform transpose operation on the last two dimensions; This is the scaling factor. 4.The short-term power load forecasting method based on graph-enhanced time-frequency dual-branch structure according to claim 1, wherein, Step S2 includes: The temporal feature representation is processed by a parallel multi-branch one-dimensional convolutional structure. The temporal feature representation is input into each branch of the one-dimensional convolutional structure to obtain the convolutional output feature of each branch. In the parallel multi-branch one-dimensional convolutional structure, each branch is set with a convolutional kernel of a different size; The convolutional output features are sequentially subjected to batch normalization, ReLU activation, and Dropout processing to obtain the preprocessed features of each branch. The preprocessed features are concatenated along the channel dimension to obtain a multi-scale joint feature map; The multi-scale joint feature map is fused and its dimensionality is compressed by a one-dimensional convolutional layer with a kernel size of 1 to obtain a fused feature with the same channel dimension as the temporal feature representation. The fused features and the temporal feature representations are fused through residual connections to obtain multi-scale enhanced features. 5.The short-term power load forecasting method based on graph enhanced time-frequency dual-branch structure according to claim 1, wherein, Step S3 includes: The multi-scale enhanced features are input into a temporal coding network based on the iTransformer architecture for processing to obtain preliminary temporal prediction results; The multi-scale enhanced features are input into the frequency domain modeling network for processing to obtain preliminary frequency domain prediction results. 6.The short-term power load forecasting method based on graph-enhanced time-frequency dual-branch structure according to claim 1, wherein, The graph-based variable correlation enhancement in step S4 specifically includes: Average pooling is performed on the feature tensors corresponding to the preliminary prediction results in the time domain and the preliminary prediction results in the frequency domain respectively to obtain the variable-level feature representations for each variable. The variable-level feature representation is subjected to L2 norm normalization, and the cosine similarity between variables is calculated based on the normalized variable-level feature representation to generate an initial similarity matrix. The initial similarity matrix is ​​sparsified to obtain a sparse similarity matrix. An adjacency matrix is ​​constructed by adding an identity matrix to the sparse similarity matrix. Row normalization is then performed on the adjacency matrix using the degree matrix to obtain a normalized adjacency matrix. Based on the normalized adjacency matrix, graph convolution feature aggregation is performed on the feature tensors corresponding to the preliminary prediction results in the time domain and the feature tensors corresponding to the preliminary prediction results in the frequency domain, respectively, to obtain graph convolution time domain features and graph convolution frequency domain features; Nonlinear transformations are performed on the graph convolutional time-domain features and graph convolutional frequency-domain features respectively, and then fused with the corresponding preliminary prediction results through residual connections to obtain the graph-enhanced time-domain features and graph-enhanced frequency-domain features. 7.The short-term power load forecasting method based on graph enhanced time-frequency dual-branch structure according to claim 1, wherein, The adaptive weighted fusion in step S5 specifically includes: Perform a discrete Fourier transform on the standardized sequence to obtain the spectral information corresponding to the sequence. Calculate the energy ratio of the main frequency and harmonic components based on the spectral information to obtain the time-domain fusion weight and frequency-domain fusion weight. Based on the time-domain fusion weights and frequency-domain fusion weights, the graph-enhanced time-domain features and graph-enhanced frequency-domain features are weighted and summed to obtain the final power load prediction result.

8. A short-term power load forecasting system based on a graph-enhanced time-frequency dual-branch structure, characterized in that, include: The data preprocessing module (1) is used to preprocess the input power load sequence to obtain a standardized sequence; and to perform adaptive time-domain query on the standardized sequence to obtain a phase-enhanced time-series feature representation. The multi-scale feature extraction module (2) is used to perform multi-scale convolution feature extraction on the temporal feature representation to obtain multi-scale enhanced features; The dual-branch processing module (3) is used to input the multi-scale enhancement features into the time-domain branch and the frequency-domain branch respectively for processing, and obtain the preliminary prediction results in the time domain and the preliminary prediction results in the frequency domain. The graph enhancement module (4) is used to enhance the variable correlation of the time-domain preliminary prediction result and the frequency-domain preliminary prediction result based on the graph structure, so as to obtain graph-enhanced time-domain features and graph-enhanced frequency-domain features. The fusion prediction module (5) is used to adaptively weight and fuse the graph-enhanced time-domain features and graph-enhanced frequency-domain features to obtain the final power load prediction result. 9.The short-term power load forecasting system based on graph enhanced time-frequency dual-branch structure according to claim 8, wherein, The data preprocessing module (1) includes: The standardization unit (11) is used to perform Z-score standardization on all numerical features of the power load sequence and divide the standardized power load sequence into training set, validation set and test set in time order to obtain the divided power load sequence. The periodic parameter initialization unit (12) is used to initialize the periodic parameter matrix used for periodic modeling in the adaptive time domain query for the divided power load sequence. The periodic parameter matrix is ​​initialized uniformly by Xavier. The query tensor generation unit (13) is used to calculate the relative index of each time step within the corresponding period length of the period parameter matrix through modular operation, retrieve the query vector from the period parameter matrix based on the relative index, and generate a query tensor. The context information aggregation unit (14) is used to take the query tensor as the query vector, take the embedding features of the normalized sequence as the key and value, and calculate the aggregated context information through the scaling dot product attention mechanism. The temporal feature representation generation unit (15) is used to fuse the aggregated context information with the embedded features of the standardized sequence through residual connection to obtain the temporal feature representation. 10.The short-term power load forecasting system based on graph enhanced time-frequency dual-branch structure according to claim 9, wherein, The graph enhancement module (4) includes: The variable-level feature generation unit (41) is used to perform average pooling processing on the feature tensor corresponding to the preliminary prediction result in the time domain and the feature tensor corresponding to the preliminary prediction result in the frequency domain, respectively, to obtain the variable-level feature representation corresponding to each variable. The similarity calculation unit (42) is used to perform L2 norm normalization on the variable-level feature representation, calculate the cosine similarity between variables based on the normalized variable-level feature representation, and generate an initial similarity matrix; The adjacency matrix construction unit (43) is used to perform sparsification on the initial similarity matrix to obtain a sparse similarity matrix, add an identity matrix to the sparse similarity matrix to construct an adjacency matrix, and perform row normalization on the adjacency matrix through the degree matrix to obtain a normalized adjacency matrix. Graph convolution unit (44) is used to perform graph convolution feature aggregation on the feature tensor corresponding to the preliminary prediction result in the time domain and the feature tensor corresponding to the preliminary prediction result in the frequency domain, based on the normalized adjacency matrix, to obtain graph convolution time domain features and graph convolution frequency domain features. The graph enhancement feature generation unit (45) is used to perform nonlinear transformations on the graph convolution time-domain features and graph convolution frequency-domain features respectively, and then fuse them with the corresponding preliminary prediction results through residual connection to obtain the graph enhancement time-domain features and graph enhancement frequency-domain features.

Images