A cross-dimension multi-scale fusion load prediction method based on multi-user load space-time correlation

Abstract

The application belongs to the technical field of power system load prediction, and discloses a cross-dimension multi-scale fusion load prediction method based on multi-user load space-time correlation, which firstly pre-processes user load statistical data, extracts time sequence dependence and periodic characteristics in the time sequence, and introduces a channel attention mechanism to adaptively mine key variable information; then, through a multi-scale space-time fusion module combined with frequency domain analysis and a graph convolution network, deep feature interaction under different time scales and spatial levels is realized; finally, a linear projection structure is used to output load prediction results under multiple time granularities. Compared with existing methods, the application has significant advantages in capturing complex load patterns, improving model prediction accuracy and enhancing generalization ability, and is suitable for various application scenarios such as power user energy consumption management and power grid load scheduling.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power load prediction, and particularly relates to a cross-dimension multi-scale fusion load prediction method based on multi-user load space-time correlation. BACKGROUND

[0002] Power load prediction is a key technology in energy management systems, and its accuracy directly affects the efficiency of power grid dispatching and power supply reliability. With the popularity of distributed energy, electric vehicles and diversified electrical equipment, user-side load data presents characteristics of high nonlinearity, multi-scale and strong randomness, and traditional prediction methods face significant challenges. Current mainstream prediction models mainly include time series analysis methods based on statistics (such as ARIMA and exponential smoothing) and end-to-end models based on deep learning (such as LNN, TCN and Transformer). Statistical methods are computationally efficient, but they are difficult to capture complex nonlinear features in load data. Deep learning models have advantages in feature extraction, but they still have limitations in multi-scale feature fusion. For example, LSTM networks can effectively model temporal dependencies, but they are not responsive to sudden load changes. The Transformer architecture is good at capturing long-term dependencies, but it is less sensitive to high-frequency fluctuations and has high computational complexity. In addition, existing methods only rely on time dimension modeling, ignoring the inherent spatial correlation in power systems. This single-dimensional modeling approach limits the model's generalization ability in complex scenarios, especially in industrial electricity scenarios, making it difficult to capture cross-regional load propagation effects, such as power grid cascading fluctuations caused by the start and stop of a production line, further increasing the difficulty of prediction. SUMMARY

[0003] The application aims to provide a cross-dimension multi-scale fusion load prediction method based on multi-user load space-time correlation to solve the technical problem of insufficient response to sudden load and gradual changes in traditional methods.

[0004] To solve the above technical problems, the specific technical solutions of the cross-dimension multi-scale fusion load prediction method based on multi-user load space-time correlation are as follows:

[0005] A cross-dimension multi-scale fusion load prediction method based on multi-user load space-time correlation includes the following steps:

[0006] Step 1: Obtain user load statistical data set;

[0007] Step 2: Preprocess the obtained data and divide the data set into training set, validation set and test set;

[0008] Step 3: Build a load prediction model based on cross-dimension multi-scale fusion;

[0009] Step 4: Train the cross-dimension multi-scale fusion network using the training set and validation set obtained in step 2, select the model parameters with the minimum validation loss on the validation set as the optimal model parameters, and perform performance evaluation on the optimal model using the test set;

[0010] Step 5: Input the load history monitoring data of the user to be predicted into the trained load prediction model, and output the load prediction of a certain time granularity in the future.

[0011] Further, the data in step 2 is a user load statistical data set, which contains the electricity load statistical data of multiple users in a unified time range, and the data has clear timestamp labeling, forming a multi-user, multi-time point time series data structure. All data are uniformly preprocessed by the system; then the data set is divided into training set, validation set and test set according to the user division strategy.

[0012] Further, the load prediction model based on cross-dimension multi-scale fusion in step 3 includes a time scale embedding module, a channel attention module, a multi-scale space-time fusion module, and a linear projection prediction output module, each module is used to perform the following steps:

[0013] Step 3.1: Extract the time series dependence and periodic variation characteristics of the load data on different time scales through the time scale embedding module;

[0014] Step 3.2: Adaptively capture key feature channels through the channel attention module to enhance the model's attention ability to different load feature dimensions;

[0015] Step 3.3: Through the multi-scale space-time fusion module, the features are interactively fused on different time and spatial resolutions, further integrating information of different scales to enhance the overall representation ability;

[0016] Step 3.4: Input the fused deep features into the linear projection prediction output module to realize high-precision prediction of future load change trend.

[0017] Further, step 3.1 includes the following steps:

[0018] Step 3.1.1: Normalize each element of the time series of each data variable to the range of [-1, 1]; X i

[0019] Step 3.1.2: Project the normalized sequence to a dimensional space through a 1D convolutional layer to obtain a basic feature mapping;

[0020] Step 3.1.3: Superimpose position embedding to encode sequence time position information;​

[0021] Step 3.1.4: Introducing learnable global timestamp embedding to capture periodic temporal features;

[0022] Step 3.1.5: Generating embedding with unified input representation.

[0023] Further, the step 3.2 includes the following steps:

[0024] Step 3.2.1: First, a global average pooling operation is performed on the input data to obtain the overall state of the time series throughout its life cycle, obtaining a global mean vector of each variable ; At the same time, a global standard deviation operation is used to measure the fluctuation amplitude of each variable in the time dimension, thereby obtaining a standard deviation vector reflecting the degree of dynamic change ;

[0025] Step 3.2.2: The global mean vector and the standard deviation vector are respectively input into two independent 1x1 convolutional layers, and high-order channel expressions are extracted through further convolutional operations, thereby forming channel embedding features representing the importance of each variable and , then splicing and using convolutional operations to extract high-level features ;

[0026] Step 3.2.3: Normalizing the high-level features , and generating attention weight coefficients of each variable through Sigmoid activation function ; The weight coefficients are multiplied element by element with the original embedding vector to generate a weighted variable feature representation tensor , realizing the weighted adjustment in the variable dimension;

[0027] Finally, by merging the variable dimension and the channel dimension, the weighted variable feature matrix is converted into a shape of .

[0028] Further, the step 3.3 includes the following steps:

[0029] Step 3.3.1: In the multi-scale recognition layer of the multi-scale spatio-temporal fusion module, first, the fast Fourier transform is used to convert the time series representation of the embedding from the time domain to the frequency domain, decompose different frequency components, and calculate the amplitude corresponding to each frequency point.

[0030] Step 3.3.2: Select the first k frequency components As the main periodic signal source, let the model focus on the periodic component that has the most significant impact on the sequence fluctuation, filter the noise frequency, and convert the sequence into a corresponding time scale through frequency ; L ; ;

[0031] Step 3.3.3: Perform time dimension expansion on the sequence , ensure that the sub-sequence length is aligned with the time scale, and convert the completed to a multi-scale feature vector suitable for graph convolution through dimension reshaping operation ;

[0032] Step 3.3.4: In the adaptive graph convolution layer of the multi-scale spatio-temporal fusion module, receive the reconstructed sub-sequence , project the tensor corresponding to the th scale back to the variable space through a learnable linear transformation matrix i to obtain , generate a trainable parameter matrix ;

[0033] Step 3.3.5: Multiply the two parameter matrices to obtain the original association matrix, filter the negative association through the activation function, and then normalize the weight between different nodes to obtain the adaptive adjacency matrix suitable for the current scale ;

[0034] Step 3.3.6: Implement high-order correlation capture using the Mixhop strategy, first define the power set of the adjacency matrix, perform multi-order power operation on the adaptive adjacency matrix to obtain the j order correlation information between nodes, then perform graph convolution operation on each order matrix and input feature , and concatenate the output results according to the feature dimension, enhance the non-linear expression ability through the activation function, and finally perform dimension projection and non-linear mapping on the fusion features through the MLP multi-layer perceptron, project the graph convolution output back to a three-dimensional tensor ;

[0035] Step 3.3.7: In the multi-head attention mechanism module of the multi-scale spatio-temporal fusion module, focus on the feature interaction in the time scale dimension, first adjust the structure through dimension reshaping for the input tensor , and then generate attention input using learnable matrix based on the query Q , key K , and value V triple mechanism​

[0036] Step 3.3.8: Apply attention weight to the embedded sequence of each time scale to learn the relative importance in the time scale dimension, combine the normalized amplitude of each frequency component to generate scale aggregation weight, and realize dynamic integration and fusion of multi-scale features;

[0037] Step 3.3.9: Based on the amplitude of each scale corresponding to the fast Fourier transform calculation , generate scale aggregation weight , ensure that the weight sum is 1, and then combine the attention output and the scale weight to obtain the output result of the fused multi-scale feature through normalized weighted aggregation .

[0038] Further, the step 3.4 includes the following steps:

[0039] Step 3.4.1: Introduce a residual connection mechanism by element-wise adding the original input feature and the output feature after multi-scale processing;

[0040] Step 3.4.2: Convert the fused feature into a prediction result through a bilinear transformation, first define a prediction granularity function, and output the corresponding time granularity m for the first prediction task;

[0041] Step 3.4.3: Use a learnable weight matrix to map the time dimension of the fused feature from L to the prediction length T to obtain the intermediate feature , realize cross-scale conversion from the length of the historical sequence to the length of the future prediction, while keeping the feature dimension unchanged, and then through the weight matrix , map the feature dimension of from to the variable number N to obtain the variable space feature , realize the prediction conversion from abstract feature to specific variable;

[0042] Step 3.4.4: Combine the bias vector to generate the prediction result of the future time steps .

[0043] Further, the step 4 includes the following steps:

[0044] Use the training set to train the built prediction model, and use the trained model to infer the validation set, and calculate the root mean square errorRMSE and MAE Two evaluation indicators are used to verify the model performance, the model with the best performance on the validation set is selected as the best model, and the test set is used to evaluate the performance of the best model.

[0045] The multi-user load space-time correlation based cross-dimension multi-scale fusion load prediction method has the following advantages:

[0046] (1) The present application fuses cross-dimension information and multi-scale time series features on the basis of the backbone network, and significantly improves the understanding ability of the model to user load behavior through the design of multi-path fusion structure, effectively enhancing the accuracy and stability of the prediction result.

[0047] (2) The present application combines the multi-dimensional information of the spatial dimension and the time dimension of the user load time series data, and performs feature enhancement, which significantly improves the accuracy of load prediction.

[0048] (3) The present application designs a multi-scale feature extraction module, adopts a sliding window processing strategy to obtain load sequence features at multiple time granularities, improves the common perception ability of the model to short-term fluctuations and long-term trends, and enhances the recognition ability of the load periodicity and suddenness.

[0049] (4) The present application introduces an attention mechanism, adjusts the weight of important features in the fusion stage, so that the model focuses on key influencing factors, thereby suppressing the interference of redundant information and improving the robustness of load prediction in complex environments.

[0050] (5) The present application proposes a deep fusion module, which gradually fuses cross-dimension and multi-scale features in multiple stages, guides the model to construct load prediction representation from coarse to fine, and then generates a load curve with better spatio-temporal consistency and prediction accuracy.

[0051] (6) The present application introduces a module-level cascading mechanism based on residual connection, which deeply fuses features of each dimension while maintaining the integrity of information transmission, and performs end-to-end training through a joint loss function to optimize the overall performance of the model. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 The flowchart of the multi-user load space-time correlation based cross-dimension multi-scale fusion load prediction method of the present application;

[0053] Figure 2 is a user load prediction model structure diagram;

[0054] Figure 3 is a time scale embedding module structure diagram;

[0055] Figure 4 is a channel attention module structure diagram;

[0056] Figure 5 This is a structural diagram of a multi-scale spatiotemporal fusion module;

[0057] Figure 6 It is an adaptive graph convolutional layer structure diagram;

[0058] Figure 7 This is a structural diagram of the multi-head attention mechanism module. Detailed Implementation

[0059] To better understand the purpose, structure, and function of this invention, the following description, in conjunction with the accompanying drawings, provides a more detailed account of a cross-dimensional, multi-scale fusion load prediction method based on the spatiotemporal correlation of multi-user loads.

[0060] like Figure 1 As shown, the present invention provides a cross-dimensional, multi-scale fusion load forecasting method based on the spatiotemporal correlation of multi-user load, comprising the following steps:

[0061] Step 1: Obtain user load statistics data set;

[0062] Step 2: Preprocess the acquired data by dividing the dataset into training, validation, and test sets;

[0063] The data described is a user load statistics dataset, which includes the following time-series monitoring data: Regarding electricity load, statistical data on the electricity load of 320 users within a unified time range were collected. The data has clear timestamp annotations, forming a typical multi-user, multi-time-point time-series data structure. All data underwent a systematic preprocessing process: first, data quality control was performed, including format standardization and handling of missing values; then, feature normalization was implemented to improve the stability and efficiency of model training; finally, time-correlated model input samples were generated using time-series sample construction methods such as sliding windows. This dataset is characterized by its clear data structure and consistent time granularity, making it suitable for multi-user load prediction tasks. In terms of data partitioning, the dataset is divided into training, validation, and test sets by user, ensuring a more objective evaluation of the model's generalization ability when facing unknown user data. This data organization method effectively supports the cross-dimensional, multi-scale fusion load prediction modeling requirements based on the spatiotemporal correlation of multi-user load.

[0064] The time-series monitoring data sample a is... ,in m It is the number of variables, the first i The time series of the variables are ,in For time step.

[0065] Step 3: Build a load forecasting model based on cross-dimensional and multi-scale fusion;

[0066] As shown in Figure 2 , the load prediction model based on cross-dimension multi-scale fusion includes a time scale embedding module, a channel attention module, a multi-scale space-time fusion module, and a linear projection prediction output module, each of which is used to perform the following steps:

[0067] Step 3.1: As shown in Figure 3 , the time series dependence and periodic variation characteristics of the load data on different time scales are extracted by the time scale embedding module.

[0068] Step 3.1.1: The time series of each data variable is normalized to the range of [-1, 1], and the specific expression method is as follows: X i

[0069]

[0070] wherein, L denotes the length of the time series, t denotes the initial position of the prediction window.

[0071] Step 3.1.2: Through the 1D convolution layer, the normalized sequence is projected into dimensional space to obtain the basic feature mapping.

[0072] Step 3.1.3: Superimpose position embedding to encode the sequence time position information, and the specific expression is as follows:

[0073]

[0074] wherein, is a mathematical symbol, representing the real number set, denotes the position, i.e. the time step in the time series.

[0075] Step 3.1.4: Introduce a learnable global timestamp embedding to capture periodic time features such as minutes / hours.

[0076] Step 3.1.5: Use a unified input representation to generate embedding , wherein s , are the variable dimension and channel dimension, respectively, and the specific expression is as follows:

[0077]

[0078] wherein, is an adjustable hyperparameter used to control the contribution of the basic feature mapping,​ 1×1 convolution operation.

[0079] Step 3.2: As Figure 4 As shown, the channel attention module adaptively captures key feature channels, enhancing the model's ability to focus on different load feature dimensions.

[0080] Step 3.2.1: First, perform global average pooling on the input data to obtain the overall state of the time series over its entire lifecycle, and obtain the global mean vector for each variable. Simultaneously, a global standard deviation operation is used to measure the fluctuation range of each variable over time, thereby obtaining a standard deviation vector that reflects the degree of its dynamic change. These two global statistical features characterize the sequence properties from different perspectives. Their combination preserves the overall trend while enhancing the responsiveness to local temporal changes. The specific expression is as follows:

[0081]

[0082] in, , These are the average pooling feature and the standard deviation feature, respectively. and These are global average pooling and global standard deviation operations, respectively.

[0083] Step 3.2.2: Convert the global mean vector and standard deviation vector The inputs are fed into two independent 1×1 convolutional layers, and higher-order channel representations are extracted through further convolution operations, thus forming channel embedding features that represent the importance of each variable. and Then, the data is concatenated and convolutional operations are used to extract high-level features. The specific expression is as follows:

[0084]

[0085] in, It is an activation function. It's a concatenation operation; the feature variables are obtained through concatenation. .

[0086] Step 3.2.3: High-level features Perform normalization processing, and through Sigmoid The activation function generates attention weight coefficients for each variable. The weighting coefficients are multiplied element-wise with the original embedding vector to generate a weighted variable feature representation tensor. , which realizes the weighted adjustment in the variable dimension, significantly enhancing the model's focus on key variables, expressed as follows:

[0087]

[0088] wherein, is the activation function, represents the batch normalization operation, is the attention weight.

[0089] Finally, by merging the variable dimension and the channel dimension, the weighted variable feature matrix is converted into a shape of .

[0090] Step 3.3: As shown in Figure 5 , the multi-scale spatio-temporal fusion module interacts and fuses features at different temporal and spatial resolutions, further integrating information at different scales to enhance overall representation ability.

[0091] The proposed multi-scale spatio-temporal fusion module is used to perform the following steps:

[0092] Step 3.3.1: In the multi-scale identification layer of the multi-scale spatio-temporal fusion module, first, the embedded time series representation is converted from the time domain to the frequency domain using the fast Fourier transform, decomposing different frequency components, and calculating the amplitude corresponding to each frequency point, expressed as follows:

[0093]

[0094] wherein, represents performing the fast Fourier transform on the input time series data, represents calculating the amplitude corresponding to each frequency component in the frequency domain, represents the index of the channel dimension, and represent the real part and the imaginary part of the extracted channel, respectively. The function performs averaging operations in the channel dimension, and the vector represents the calculated amplitude of each frequency. is the k frequency component, j represents the imaginary unit .

[0095] Step 3.3.2: Select the top k frequency components with the largest amplitude values as the main periodic signal source, allowing the model to focus on the periodic components that have the most significant impact on sequence fluctuations, filtering out noise frequencies. Through the frequency with the inverse relationship of time series length L is converted to the corresponding time scale , the specific expression is as follows:

[0096]

[0097] wherein, the first k frequency index with the largest amplitude is selected. is the frequency converted into the corresponding time scale.

[0098] Step 3.3.3: In order to unify the sub-sequences of different time scales to input the graph convolution module, the time dimension expansion is performed on the sequence to ensure that the sub-sequence length is aligned with the time scale, and to avoid dimensional misplacement during convolution. Through the dimension reshaping operation, the completed is converted into a multi-scale feature vector adapted to graph convolution, the specific expression is as follows:

[0099]

[0100] wherein, is an integer multiple of . is used to zero-expand the time series in the time dimension, so as to be compatible with , denotes the reconstructed time series based on the time scale .

[0101] Step 3.3.4: As shown in Figure 6 , in the adaptive graph convolution layer of the multi-scale spatio-temporal fusion module, the reconstructed sub-sequence is received, and the tensor corresponding to the i th scale is projected back to the variable space through a learnable linear transformation matrix to obtain , and a trainable parameter matrix is generated, wherein N is the number of variables, h is the hidden dimension, and the specific expression is as follows:

[0102]

[0103] wherein, denotes the variable space projection of the i th time scale, is a learnable weight matrix.

[0104] Step 3.3.5: Combine these two parameter matrices Multiplying the results yields the original association matrix. Negative associations are then filtered out using an activation function. Finally, the weights between different nodes are normalized to obtain an adaptive adjacency matrix suited to the current scale. The specific expression is as follows:

[0105]

[0106] in, ReLU It is an activation function that ensures all negative values ​​become zero. SoftMax The function is used to normalize the weights of different nodes.

[0107] Step 3.3.6: Employ the Mixhop strategy to capture higher-order relevance. First, define the power set of the adjacency matrix, and then apply the adaptive adjacency matrix... Perform multi-order power operations to obtain the relationships between nodes. j The order-related information is then analyzed. Subsequently, the matrices of each order are compared with the input features. Graph convolution operations are performed, and the output is concatenated according to the feature dimensions. An activation function is then used to enhance the non-linear expressive power. Finally, through... MLP The multilayer perceptron performs dimensionality projection and nonlinear mapping on the fused features, and outputs graph convolution. Projecting back to 3D tensor This improves the model's ability to capture complex dependencies between variables, as specifically expressed below:

[0108]

[0109] Among them, hyperparameters p It is the set of powers of the adjacency matrix. Represents the learned adjacency matrix Multiply by itself j Second-rate, This indicates concatenating neighbor features at each level column by column. It is an activation function. This represents dimensional projection and nonlinear mapping.

[0110] Step 3.3.7: As Figure 7 As shown, in the multi-head attention mechanism module of the multi-scale spatiotemporal fusion module, the focus is on the feature interactions in the time scale dimension. First, the input tensor... The structure is adjusted by reshaping the dimensions, and then based on the query. Q ,key K ,value V The triplet mechanism, using learnable matrices Generate attention input, specifically expressed as follows:

[0111]

[0112] wherein, B is the batch size, denotes a reshaping function, is a mapping matrix.

[0113] Step 3.3.8: Apply attention weighting to the embedded sequences of each time scale, learn the relative importance in the time scale dimension, further combine the normalized amplitudes of each frequency component to generate scale aggregation weights, realize dynamic integration and fusion of multi-scale features, and improve the comprehensive extraction ability of the model to periodic and asynchronous characteristics. The input tensor is reshaped to , and the specific expression is as follows:

[0114]

[0115] wherein, SoftMax is a function for calculating attention weights and aggregating features, is the dimension of the key vector k in multi-head attention to avoid gradient disappearance when the dimension is high. is a multi-head self-attention function that concatenates multi-head attention outputs and projects back to the original feature dimension, c is the number of attention heads, is the output projection matrix.

[0116] Step 3.3.9: Based on the amplitudes of each scale calculated by the fast Fourier transform , generate scale aggregation weights , ensure that the sum of the weights is 1, and then combine the attention output and the scale weight to obtain the output result of the fused multi-scale feature by normalized weighted aggregation, and the specific expression is as follows:

[0117]

[0118] wherein, is the amplitude normalization value corresponding to the i th frequency. denotes an amplitude normalization operation for weighted fusion of features of different time scales. denotes normalization of the sample feature layer.

[0119] Step 3.4: Input the fused deep features into the linear projection prediction output module to realize high-precision prediction of future load change trends.

[0120] The proposed linear projection prediction output module is used to perform the following steps:

[0121] Step 3.4.1: The residual connection mechanism is introduced by element-wise adding the original input features and the output features after multi-scale processing, effectively alleviating the gradient vanishing problem in deep networks, while improving the stability and expressive power of model training. The specific expression is as follows:

[0122]

[0123] wherein, l is the network layer index.

[0124] Step 3.4.2: The fusion features are converted into prediction results through bilinear transformation. First, define the prediction granularity function, which outputs the corresponding time granularity m for the th prediction task. The specific expression is as follows:

[0125]

[0126] wherein, is used to support multi-time granularity output, is the m th prediction granularity (which includes hourly levels such as 1 hour, 3 hours).

[0127] Step 3.4.3: Use the learnable weight matrix to map the time dimension of the fusion features from L to the prediction length T to obtain the intermediate features , realizing the cross-scale conversion of historical sequence length to future prediction length while keeping the feature dimension unchanged. Then through the weight matrix , the feature dimension of is mapped from to the variable number N to obtain the variable space features , realizing the prediction conversion of abstract features to specific variables. The specific expression is as follows:

[0128]

[0129] Step 3.4.4: Combine the bias vector to generate the prediction results for the next time steps. The specific expression is as follows:

[0130]

[0131] wherein, is the bias vector, is the future load prediction value of the future time step.

[0132] Step 4: training the cross-dimension multi-scale fusion network using the training set and the validation set obtained in step 2, selecting the model parameters with the minimum validation loss on the validation set as the best model parameters, and performing performance evaluation on the best model using the test set;

[0133] In step 4, the training set is used to train the built prediction model, and the trained model is used to infer the validation set. The model performance is verified by two evaluation indexes, root mean square error RMSE and MAE , and the best model is selected as the best model. The performance of the best model is evaluated using the test set, and the specific expression is as follows:

[0134]

[0135] wherein, n is the number of test samples, and are the true value and the predicted value of the i th test sample, respectively. i

[0136] Step 5: inputting the load history monitoring data of the user to be predicted into the trained user load prediction model, and outputting the load prediction of a certain time granularity in the future.

[0137] It can be understood that the present application is described through some embodiments, and those skilled in the art know that various changes or equivalent replacements can be made to these features and embodiments without departing from the spirit and scope of the present application. In addition, under the guidance of the present application, these features and embodiments can be modified to adapt to specific conditions and materials without departing from the spirit and scope of the present application. Therefore, the present application is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of the present application are within the scope of the present application.​

Claims

1. A multi-dimensional multi-scale fusion load prediction method based on multi-user load space-time correlation, characterized in that, The method comprises the following steps: Step 1: Obtain a user load statistical data set, and pre-process the obtained data; Step 2: Build a load prediction model based on cross-dimension multi-scale fusion; the load prediction model comprises a time scale embedding module, a channel attention module, a multi-scale space-time fusion module and a linear projection prediction output module, and each module is used for performing the following steps: Step 2.1: Extract the time sequence dependence and periodic change characteristics of the load data on different time scales through the time scale embedding module; Step 2.2: Adaptively capture key feature channels through the channel attention module, and enhance the attention ability of the model to different load feature dimensions; Step 2.3: Through the multi-scale space-time fusion module, the features are interactively fused on different time and space resolutions, and the information of different scales is further integrated to enhance the overall representation ability; Step 2.3.1: In the multi-scale recognition layer of the multi-scale space-time fusion module, first, the fast Fourier transform is used to convert the embedded time sequence representation from the time domain to the frequency domain, decompose different frequency components, and calculate the amplitude corresponding to each frequency point; Step 2.3.2: Select the first value with the largest amplitude. k Frequency components As the primary source of periodic signals, the model focuses on the periodic components that have the most significant impact on sequence fluctuations, filtering out noise frequencies and using frequency... With sequence length L The inverse proportional relationship is converted to obtain the corresponding time scale. ; Step 2.3.3: On the sequence Perform temporal dimension expansion, ensuring that the sub-sequence length is aligned with the time scale, complete the converted into a multi-scale feature vector adapted to graph convolution ; Step 2.3.4: In the adaptive graph convolution layer of the multi-scale spatio-temporal fusion module, the above multi-scale feature vector is received , the tensor corresponding to the first i is projected back to the variable space by a learnable linear transformation matrix to obtain , and a trainable parameter matrix is generated;​ Step 2.3.5: Multiply the two parameter matrices to obtain the original association matrix, filter the negative association through the activation function, and normalize the weights between different nodes to obtain the adaptive adjacency matrix adapted to the current scale ; Step 2.3.6: Employ the Mixhop strategy to capture higher-order relevance. First, define the power set of the adjacency matrix, and then apply the adaptive adjacency matrix... Perform multi-order power operations to obtain the relationships between nodes. j The order correlation information is then used to correlate the matrices of each order with the input features. Perform graph convolution operations and concatenate the outputs according to feature dimensions. Enhance the nonlinear expressive power through activation functions, and finally... MLP The multilayer perceptron performs dimensionality projection and nonlinear mapping on the fused features, and outputs graph convolution. Projecting back to 3D tensor ; Step 2.3.7: In the multi-head attention mechanism module of the multi-scale spatiotemporal fusion module, focus on the feature interactions in the time scale dimension, first targeting the input tensor. The structure is adjusted by reshaping the dimensions, and then based on the query. Q ,key K ,value V The triplet mechanism, using learnable matrices Generate attention input; Step 2.3.8: Apply attention weighting to the embedded sequence of each time scale, learn the relative importance in the time scale dimension, combine the normalized amplitudes of each frequency component to generate scale aggregation weights, and realize dynamic integration and fusion of multi-scale features; Step 2.3.9: Calculate the amplitude of each scale corresponding to the fast Fourier transform , generate scale aggregation weight , ensure the weight sum is 1, and combine the attention output with the scale weight to obtain the output result of the fused multi-scale feature through normalized weighted aggregation ; Step 2.4: Input the fused deep features into the linear projection prediction output module to realize high-precision prediction of the future load change trend; Step 3: Input the load historical monitoring data of the user to be predicted into the trained load prediction model, and output the load prediction of a certain time granularity in the future.

2. The multi-user load spatio-temporal correlation based cross-dimension multi-scale fusion load forecasting method according to claim 1, characterized in that, The data in step 1 is a user load statistical data set, which contains the power load statistical data of multiple users within a unified time range, and the data has clear timestamp labeling, forming a multi-user and multi-time point time sequence data structure. All data are uniformly pre-processed by the system; then the data set is divided into a training set, a validation set and a test set according to the user division strategy.

3. The multi-user load spatio-temporal correlation based cross-dimension multi-scale fusion load forecasting method according to claim 1, characterized in that, The step 2.1 comprises the following steps: Step 2.1.1: Time series for each data variable X i Normalization is performed to normalize each element to the range [-1, 1]; Step 2.1.2: Project the normalized sequence into a feature map by a 1D convolutional layer a 256-dimensional feature space. Step 2.1.3: Superimpose position embedding to encode sequence time position information; Step 2.1.4: Introduce a learnable global timestamp embedding to capture periodic time features; Step 2.1.5: Use a unified input representation to generate embedding.

4. The multi-user load spatio-temporal correlation based cross-dimension multi-scale fusion load forecasting method according to claim 1, characterized in that, The step 2.2 comprises the following steps: Step 2.2.1: First, a global average pooling operation is performed on the input data to obtain the overall state of the time series throughout its life cycle, resulting in a global mean vector for each variable ; At the same time, a global standard deviation operation is used to measure the fluctuation amplitude of each variable in the time dimension, resulting in a standard deviation vector reflecting the degree of dynamic change ; Step 2.2.2: The global mean vector and the standard deviation vector are input into two independent 1x1 convolutional layers, respectively, to extract high-order channel representations through further convolutional operations, thereby forming channel embedding features representing the importance of each variable and Then, they are spliced and high-level features are extracted using convolutional operations ; Step 2.2.3: High-level features Perform normalization processing, and through Sigmoid The activation function generates attention weight coefficients for each variable. The weighting coefficients are multiplied element-wise with the original embedding vector to generate a weighted variable feature representation tensor. This enables weighted adjustments across variable dimensions; Finally, by merging the variable dimension with the channel dimension, the weighted variable feature matrix is transformed into a shape .

5. The multi-user load spatio-temporal correlation based cross-dimension multi-scale fusion load forecasting method according to claim 1, characterized in that, The step 2.4 comprises the following steps: Step 2.4.1: Introduce residual connection mechanism by element-wise adding the original input features with the multi-scale processed output features ​ Step 2.4.2: Convert the fused features into prediction results by a bilinear transformation, first define a prediction granularity function, for the i-th prediction task output corresponding time granularity m . ; Step 2.4.3: Utilizing the learnable weight matrix , the time dimension of the fused feature is mapped from L to the prediction length T to obtain the intermediate feature , realizing the cross-scale conversion from the history sequence length to the future prediction length while keeping the feature dimension unchanged, and then the feature dimension of is mapped from to the variable number N by the weight matrix to obtain the variable space feature , realizing the prediction conversion from the abstract feature to the specific variable; Step 2.4.4: Incorporate bias vector , generate prediction for future time steps .

6. The multi-user load spatio-temporal correlation based cross-dimension multiscale fusion load forecasting method according to claim 1, characterized in that, The step 3 comprises the following steps: The built prediction model is trained using the training set, and the trained model is used to infer the validation set, and the model performance is verified by the root mean square error RMSE and MAE Two evaluation indexes, and the best model is selected as the best model, and the performance of the best model is evaluated using the test set.

Images