An extreme weather-based power load prediction method based on spatio-temporal graph convolution network

Abstract

The application provides a kind of power load prediction method based on space-time graph convolution network under extreme weather. Including: collecting the historical load data of multiple load nodes in power system and regional meteorological element data, and through the combination of model interpretation and regression analysis, the key meteorological features that significantly affect the load are screened to construct the meteorological feature vector. Multivariate empirical mode decomposition and singular value decomposition are used to reconstruct the load data at multiple scales, and smooth and effective load feature tensors are extracted. On this basis, a graph network structure is constructed based on the physical connection relationship of the nodes, and the load and meteorological features are integrated in the time dimension to form the node time series features. The load is predicted using the space-time graph convolution network model. The application can effectively model the space-time dependence of load data, enhance the prediction accuracy of load changes under extreme weather conditions, and has good robustness and generalization ability.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power load prediction, and particularly relates to a power load prediction method based on a space-time graph convolution network under extreme weather. BACKGROUND

[0002] When an extreme weather event occurs, such as high temperature or severe cold, the power load will show a trend of rapid increase, which poses new challenges to short-term power load prediction. People's daily activities and meteorological factors bring great volatility and uncertainty to short-term power load, and these external factors have an important influence on the accuracy of power load prediction. Power load is influenced by factors such as season, climate, holiday system, region, economic policy, and presents complex changes. However, the existing short-term load prediction method has limited prediction ability for the nonlinear and dynamic changing load characteristics under extreme weather. Accurate and stable short-term load prediction is crucial for the daily operation of the power system and is the basis for building a new power system. It can provide decision-making basis for solving the imbalance between power supply and demand. At present, power load prediction methods mainly include related methods based on traditional mathematical statistics and related machine learning methods based on data driving. The related methods based on traditional mathematical statistics include time series analysis method of autoregressive integrated moving average, multivariate linear regression analysis method and exponential smoothing method. Although the statistical method has a simple model structure and fast calculation speed, it performs poorly when predicting load sequences with large fluctuations or insufficient periodicity.

[0003] Deep learning algorithms such as convolutional neural network (CNN), long short-term memory network (LSTM) and bidirectional long short-term memory network (BiLSTM) are widely used in load prediction and show good performance. However, these methods generally only consider the features of time series data, ignoring the spatial correlation between nodes in the power system, and are difficult to fully capture the complex spatio-temporal dependence of load changes under extreme weather. In addition, the power load data under extreme weather conditions is highly volatile and non-stationary, and direct application of deep learning models often fails to achieve satisfactory prediction accuracy.

[0004] In the prior art, Chinese patent CN118428544A discloses a method for periodic multi-node power load prediction and related devices, comprising: obtaining historical load and influencing factor data of a to-be-predicted load node and adjacent nodes, and constructing a multi-node power load original data set; then preprocessing to obtain a multi-node load sample data set, screening influencing factors from the multi-node load sample data set, and constructing a final multi-node load sample data set; then dividing according to working days and holidays to construct a periodic multi-node load sample data set; then fusing a graph convolution network and a bidirectional gated recurrent unit to build an initial periodic multi-node power load prediction model; and finally optimizing the hyperparameters of the initial periodic multi-node power load prediction model based on the data set to determine the final periodic multi-node power load prediction model.

[0005] However, this method has the following limitations: 1. The influencing factor screening is only based on the correlation coefficient metric, ignoring the possible nonlinear and causal relationship between meteorological factors and load changes, and cannot ensure that the selected features still have sufficient explanatory power under extreme conditions; 2. Although a graph convolution network is used to capture the spatial structure, the spatial model is connected in series with a bidirectional gated recurrent unit, and the spatial and temporal features are processed separately, making it difficult to realize joint spatio-temporal feature fusion under a multi-node, full-topology structure, especially when the physical connectivity between load nodes changes dynamically.

[0006] Therefore, there is still a lack of a short-term power load prediction method in the prior art that can effectively identify nonlinear, causal key meteorological features under extreme weather conditions, realize multi-node full-topology spatio-temporal feature deep fusion on the real topology of the power grid, and has sufficient robustness to the strong volatility and multi-scale dynamic changes of the load sequence. SUMMARY

[0007] The purpose of the present application is to overcome the defects of the prior art and provide a power load prediction method based on a spatio-temporal graph convolution network under extreme weather conditions.

[0008] The purpose of the present application can be achieved by the following technical solutions:

[0009] In one aspect, the present application provides a power load prediction method based on a spatio-temporal graph convolution network under extreme weather conditions, comprising the following features:

[0010] Collecting historical load data and regional meteorological element data of multiple load nodes in a power system;

[0011] Based on the regional meteorological element data, calculating the feature contribution value of each meteorological feature by combining model interpretation and regression analysis, and selecting several key meteorological features to construct a meteorological feature vector;

[0012] Multivariate empirical mode decomposition is performed on historical load data of multiple load nodes in the power system to extract multi-order intrinsic mode function components. Singular decomposition is then performed on each order mode component to construct a low-rank approximate embedding matrix and obtain the load feature tensor after multi-scale smooth reconstruction.

[0013] Based on the physical connection relationship of each load node in the power system, a graph network structure and adjacency matrix are constructed. At each time slice, the load feature tensor and meteorological feature vector of the corresponding node are concatenated to form the time-series feature representation of the graph network node.

[0014] Based on the constructed graph network structure and adjacency matrix, a power load prediction model based on spatiotemporal graph convolutional network is built, and power load prediction is realized based on the constructed power load prediction model.

[0015] Furthermore, the load node refers to each substation node, the historical load data refers to the active power value of the power grid measured for each corresponding substation node within a preset sampling period, and the regional meteorological element data refers to multiple meteorological characteristics measured in the same sampling period for the area where the corresponding substation is located. The meteorological characteristics include atmospheric temperature, dew point temperature, atmospheric pressure, relative humidity, wind speed, precipitation, and sunshine duration.

[0016] Furthermore, the calculation of the feature contribution value of each meteorological feature based on regional meteorological element data, through a combination of model interpretation and regression analysis, specifically includes:

[0017] Based on the regional meteorological element data and the corresponding time-time load data, a multiple linear regression model is constructed to obtain the regression coefficients of each meteorological characteristic. ;

[0018] Based on the regional meteorological data and existing load forecasting models, the model interpretation contribution value of each meteorological feature is calculated using a model interpretation algorithm. ;

[0019] Based on the regression coefficients and model interpretation contributions of each meteorological feature, calculate the characteristic contribution of each meteorological feature to the load data:

[0020]

[0021] in, For the first i The characteristic contribution value of each meteorological feature , The first i The regression coefficients of each meteorological feature and their contribution to the model explanation;

[0022] For all meteorological characteristics Sort the values ​​from largest to smallest and select the first few. kConstruct a meteorological feature vector based on the corresponding meteorological features. .

[0023] Furthermore, based on the regional meteorological element data and the corresponding time-based load data, a multiple linear regression model is constructed to obtain the regression coefficients of each meteorological characteristic. Specifically, it includes:

[0024] At each sampling time t Load increment As the dependent variable, the meteorological characteristic vector at the same time is used. Establish a multiple linear regression model with as the independent variable:

[0025]

[0026] in, , These are the time points in the historical load data. t With time t-1 The load value, m The total number of meteorological features. For the first i The regression coefficient of a meteorological feature represents the degree of linear influence of that meteorological feature on the load increment; Indicates the first i A meteorological feature at time t Sampled value, For the error term, For constant terms;

[0027] The regression coefficients of each meteorological characteristic were obtained by solving the least squares method. .

[0028] Furthermore, based on the regional meteorological element data and existing load forecasting models, the model interpretation contribution value of each meteorological feature is calculated using a model interpretation algorithm. Specifically, it includes:

[0029] For the i A meteorological feature, defined as the set of meteorological features. For any feature not included i subset of The input vector is divided into subsets. S meteorological feature vector and containing subsets S and the features to be measured i meteorological feature vector Calculate the first based on the existing load forecasting model i Model explanation contribution value of each meteorological feature :

[0030]

[0031] in, Representing a subset S The factorial of the number of meteorological features contained therein Indicates from the remainder The factorial of the number obtained after removing the feature to be measured from the meteorological features. Represents the total number of meteorological features m factorial, , These represent the input meteorological feature vectors respectively. and The output of the current load forecasting model.

[0032] Furthermore, the step of performing multivariate empirical mode decomposition on historical load data from multiple load nodes in the power system, extracting multi-order intrinsic mode function components, and performing singular decomposition on each order mode component specifically includes:

[0033] In the power system N The historical load sequences of each load node are obtained by applying multivariate empirical mode decomposition. K A sequence of intrinsic mode functions ,in, Indicates load node j At any moment t The Bar modal components;

[0034] For each modal component of each load node With the length of the sliding window L Construct the Hankel embedding matrix:

[0035]

[0036] in, Indicates load node j The bar mode components Constructed after time-delay embedding matrix; This describes an operation that transforms a one-dimensional time series into a two-dimensional matrix with a time structure, such that each antidiagonal element in the matrix is ​​identical. This represents the total number of sampling time steps in the historical load data. The length of the sliding window;

[0037] Hankel embedding matrix for each load node Perform singular value decomposition to obtain the singular value diagonal matrix. :

[0038]

[0039] wherein, is a left singular vector matrix, is a singular value matrix, is a right singular vector matrix, and T is a transpose.

[0040] Further, the constructing a low-rank approximation embedding matrix to obtain a multi-scale smooth reconstructed load feature tensor specifically comprises:

[0041] retaining the first Kmax singular values of the singular value diagonal matrix and corresponding left and right singular vectors r , and constructing a low-rank approximation embedding matrix:

[0042]

[0043] wherein, denotes a low-rank approximation embedding matrix of the load node j on the m-th modal component, is a preset constant, are the Kth singular value, left singular vector and right singular vector of the load node on the m-th modal component, respectively; j p and reconstructing into a load reconstruction signal of the load node

[0044] on the m-th modal component through inverse Hankel operation: , to obtain a multi-scale smooth reconstructed load feature tensor: j k

[0045]

[0046] wherein, denotes a load feature tensor of the load node j at time point t , denotes a load reconstruction signal of the load node j on the m-th modal component at time point k . t

[0047] Further, the power load prediction model based on the space-time graph convolutional network comprises:

[0048] a spatial graph convolution module composed of multiple layers of stacked graph convolution layers, each graph convolution layer receiving a time slice​​​​​​​t The node feature matrix and normalized adjacency matrix are taken as input, and spatial features are output through graph convolution operation;

[0049] The temporal convolution module includes multiple layers of causal one-dimensional convolution. Each layer of causal one-dimensional convolution performs convolution on the node features of the same load node and outputs temporal features.

[0050] The spatiotemporal fusion module, including a feature concatenation layer and a fusion mapping layer, is used to fuse the spatial features output by the spatial graph convolution module and the temporal features output by the temporal convolution module and map them to the prediction dimension to generate spatiotemporal features.

[0051] The output prediction submodule is used to map the spatiotemporal characteristics of each load node after fusion to the load prediction value of the next time period through one or more fully connected layers.

[0052] Furthermore, the construction of the graph network structure and adjacency matrix based on the physical connection relationships of each load node in the power system specifically includes:

[0053] Based on the meteorological feature vectors and load feature tensors of each load node, and their physical connections, a graph network structure is constructed for multiple historical time points, specifically including:

[0054] All load nodes in the power system are formed into a node set. V And form an edge set based on the physical connectivity of the transmission and distribution lines between each load node. E The node features of the graph network structure at each time step are as follows:

[0055]

[0056] in, The nodes of the graph network structure at time t eigenvectors, For load nodes j At any moment t The load characteristic tensor on, For load nodes j At any moment t Meteorological feature vectors on This represents a vector concatenation operation;

[0057] The edge weights of the graph network structure at each time step are determined by the physical properties of the power transmission and distribution lines between the load nodes, and a normalized adjacency matrix with self-loops is constructed. :

[0058]

[0059]

[0060]

[0061] wherein, denotes a node with an edge weight, is a unit matrix, N denotes the number of load nodes, is a degree matrix, whose diagonal elements ;

[0062] With a node set V, a normalized adjacency matrix and node features , define the graph network structure at each time , obtain a set of graph network structures .

[0063] Further, based on the constructed graph network structure and adjacency matrix, a power load prediction model based on a spatio-temporal graph convolution network is constructed, specifically including:

[0064] For each historical time t , the normalized adjacency matrix in the graph network structure and the node feature matrix are input into the spatial graph convolution module of the power load prediction model, the graph convolution operation is performed, the neighbor features are aggregated, and the spatial features of each node at time t are obtained:

[0065] The spatial features of each node from time to t are collected and input into the time convolution module, multi-layer causal one-dimensional convolution or dilated convolution is performed, short-term mutations and long-term trends between nodes are captured, and the time sequence features of each node are output:

[0066] The spatial features and the time sequence features are spliced in the feature dimension, and are mapped to a unified dimension through the spatio-temporal fusion module, to generate node-level spatio-temporal features;

[0067] The spatio-temporal features are input into the output prediction submodule to generate the load prediction value of each node at the next time;

[0068] The real load values of each node at the next time are taken as labels, the mean square error loss is calculated, and the parameters of the power load prediction model are iteratively updated according to the loss using back propagation and an optimizer, to realize the training of the power load prediction model.

[0069] Compared with the prior art, the present application has the following advantages:

[0070] (1) In the prior art, the screening of meteorological influencing factors only relies on the Pearson correlation coefficient, that is, the influencing factors are selected by comparing the linear correlation of each meteorological variable with the power load, which can only reflect the linear relationship between variables and cannot depict the nonlinear contribution of meteorological factors to load changes, nor can it identify the true causal driving, and it is easy to mistakenly select factors that are only accidentally related or affected by other potential variables as key features; in the feature screening stage, a multivariate linear regression model of regional meteorological elements and load increment is first constructed, and the regression coefficient of each meteorological factor is obtained by least squares method , to quantify the linear causal strength of the factor to the load change; then the marginal contribution value of each meteorological factor is calculated based on the existing load prediction model using model explanation algorithm , to comprehensively capture its influence on the model output in a nonlinear environment; finally, the two are multiplied to obtain the comprehensive feature contribution degree , and the top k meteorological features with the most driving effect are selected according to this index. In this way, the invention not only eliminates variables that are only accidentally linearly related to load changes but lack real causal action, but also retains meteorological factors that have significant marginal effects in complex nonlinear scenarios, thereby significantly improving the explanatory ability of the selected features to power load fluctuations under extreme weather conditions and enhancing the robustness and prediction accuracy of the load prediction model.

[0071] (2) The prior art usually directly uses the historical load sequence of each load node as the model input, ignoring the strong noise, high-frequency jitter and non-stationary characteristics of the load time series under extreme weather conditions. When there are sudden peaks or periodic and aperiodic components in the sequence, directly feeding into the deep learning network not only increases the difficulty of model fitting, but also easily leads to overfitting or insufficient response to dramatic fluctuations, thereby reducing the prediction accuracy and robustness. Therefore, the invention first applies multivariate empirical mode decomposition to the historical load data of multiple load nodes at the same time, decomposes the original signal into several intrinsic mode functions (IMF) and residual terms, thereby separating the trend, periodic and noise components of different frequency bands in the load sequence. Each IMF reflects the dominant fluctuation pattern on its corresponding time scale, eliminating the modal aliasing problem between different modes. Then, each IMF is time-delay embedded and a Hankel matrix is constructed, and the first r main singular values and corresponding singular vectors are extracted by singular value decomposition, completing the low-rank approximation reconstruction. This step not only removes high-frequency noise, but also retains the most representative dynamic features in each mode, making the reconstructed signal smooth and rich in multi-scale information. Finally, the reconstructed values of each node on each mode are arranged in chronological tThe load characteristic tensor is spliced, and the load change law of the node at different time scales is embodied. Through the above processing, the original load sequence with non-stationary and serious noise pollution is converted into a smooth feature tensor containing multi-scale physical meaning modal components, which significantly reduces the non-stationarity and noise interference of the data, provides more stable and more discriminative inputs for the spatial and temporal convolution modules of the spatio-temporal graph convolution network, and significantly enhances the capturing ability of the model to sudden peak values and multi-frequency fluctuations under extreme weather conditions, thereby improving the accuracy and robustness of short-term power load prediction.

[0072] (3) In the prior art, only the spatial convolution is simply connected with the time series model (such as LSTM or GRU), which leads to the fact that the spatial features and the time features are independent and interact insufficiently, and it is difficult to simultaneously capture the instantaneous spatial diffusion effect and the historical dynamic evolution trend of the power load under extreme weather conditions. In view of this defect, the present application designs a multi-layer stacked graph convolution layer under the same framework, deeply fuses the node features and the network topology of each time slice, truly reflects the electrical coupling between different substations through power transmission and distribution lines, and automatically aggregates the neighborhood information, so as to obtain more accurate and generalizable spatial representation. Then, the multi-layer causal one-dimensional convolution or dilated convolution module processes these spatial representations in parallel along the time dimension, which can not only capture short-term sudden load peaks, but also perceive multi-scale long-term fluctuation trends, making up for the limitations of RNN type algorithms in parallel computing and long-range dependence. Subsequently, the spatio-temporal fusion module splices and maps the outputs of the above two parts in the feature dimension to a unified dimension, so that the model can utilize the information of spatial topological association and temporal dynamic evolution at the same time, and generate real node-level spatio-temporal features. Finally, through the full connection submodule, these features rich in spatio-temporal interaction information are mapped to the load prediction value at the next moment, and the parameters of the graph convolution kernel, the time convolution kernel and the fusion mapping layer are continuously adjusted through back propagation in the end-to-end training process. The structure design of the present application effectively solves the problem of insufficient coupling of space and time in the prior art, and can more sensitively capture the load infection effect and mutation evolution law across nodes under extreme weather conditions, thereby significantly improving the accuracy and robustness of short-term power load prediction. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1 The power load prediction method flowchart of the embodiment of the present application;

[0074] Figure 2 The power load prediction model schematic diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0075] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of the present application.

[0076] Embodiment 1

[0077] The embodiment provides a power load prediction method based on a space-time graph convolution network in extreme weather, as shown in Figure 1 The method comprises the following features:

[0078] Step S1: collecting historical load data of a plurality of load nodes in a power system and regional meteorological element data; wherein the load nodes are each transformer substation node, the historical load data are active power values measured at each transformer substation node in a preset sampling period, and the regional meteorological element data are a plurality of meteorological characteristics measured in the same sampling period in a region where the transformer substation is located, the meteorological characteristics including atmospheric temperature, dew point temperature, atmospheric pressure, relative humidity, wind speed, precipitation, and sunshine duration.

[0079] The collected historical load data and regional meteorological element data are preprocessed, including: using linear interpolation or moving average method to complete the data with missing values, using box plot method or Z-score method to detect and eliminate the data with abnormal values; normalizing each dimension feature according to the maximum and minimum value normalization or Z-score standardization method to eliminate the dimension influence and improve the model training stability; and aligning the load data and the meteorological data according to a unified sampling period and time stamp to ensure the consistency of the input features on the time axis.

[0080] Step S2: based on the regional meteorological element data, calculating feature contribution values of each meteorological characteristic by combining model interpretation and regression analysis, and selecting a plurality of key meteorological characteristics to construct a meteorological characteristic vector, specifically including:

[0081] Based on the regional meteorological element data and the load data at the corresponding moment, a multiple linear regression model is constructed to obtain regression coefficients of each meteorological characteristic , specifically including:

[0082] Taking the load increment t of each sampling moment as the dependent variable and the meteorological characteristic vector at the same moment as the independent variable, a multiple linear regression model is established:

[0083]

[0084] wherein, , are the load values at time t and time t-1 in historical load data respectively, m is the total number of meteorological features, is the regression coefficient of the i meteorological feature, representing the degree of linear influence of the meteorological feature on the load increment; represents the sampling value of the i meteorological feature at time t , is the error term, is the constant term;

[0085] The regression coefficients of each meteorological feature are obtained by least squares method. .

[0086] Based on regional meteorological element data and existing load prediction models, the model explanation contribution value of each meteorological feature is calculated using model explanation algorithm , specifically including:

[0087] For the i meteorological feature, the meteorological feature set is defined as For any subset i not containing feature , the input vector is divided into a meteorological feature vector containing subset S and a meteorological feature vector containing subset and the to-be-tested feature S , and the model explanation contribution value of the i meteorological feature is calculated according to the existing load prediction model : i

[0088]

[0089] wherein, represents the factorial of the number of meteorological features contained in subset S , represents the factorial of the number obtained by removing the to-be-tested feature from the remaining meteorological features, represents the factorial of the total number of meteorological features m , , represent the outputs of the existing load prediction model when the input meteorological feature vector is and respectively.

[0090] ​​Based on the regression coefficients and model interpretation contributions of each meteorological feature, calculate the characteristic contribution of each meteorological feature to the load data:

[0091]

[0092] in, For the first i The characteristic contribution value of each meteorological feature , The first i The regression coefficients of each meteorological feature and their contribution to the model explanation;

[0093] For all meteorological characteristics Sort the values ​​from largest to smallest and select the first few. k Construct a meteorological feature vector based on the corresponding meteorological features. .

[0094] Through step S2 above, this invention introduces an organic combination of linear regression analysis and model interpretation methods in the meteorological feature screening process. It comprehensively considers the linear causal influence of various meteorological factors on load changes and the marginal contributions in the nonlinear model, thereby constructing a more representative feature contribution evaluation index. Based on this, the most influential key meteorological features are selected. This technique effectively compensates for the nonlinear features and spurious correlations ignored by traditional methods that rely solely on correlation metrics, eliminates noisy variables and redundant features, and improves the scientific rigor and interpretability of feature selection. Technically, this method significantly enhances the model's ability to perceive and respond to meteorological driving factors in the face of extreme weather events, improves the stability, robustness, and overall accuracy of power load forecasting, and provides a reliable data foundation for efficient training and accurate inference of subsequent models.

[0095] Step S3: Perform multivariate empirical mode decomposition on historical load data from multiple load nodes in the power system, extract multi-order intrinsic mode function components, and perform singular decomposition on each order mode component to construct a low-rank approximate embedding matrix, obtaining the load feature tensor after multi-scale smooth reconstruction. Specifically, this includes:

[0096] In the power system N The historical load sequences of each load node are obtained by applying multivariate empirical mode decomposition. K A sequence of intrinsic mode functions ,in, Indicates load node j At any moment t The Bar modal components;

[0097] For each modal component of each load node With the length of the sliding window L Construct the Hankel embedding matrix:

[0098]

[0099] in, Indicates load node j The bar mode components Constructed after time-delay embedding matrix; This describes an operation that transforms a one-dimensional time series into a two-dimensional matrix with a time structure, such that each antidiagonal element in the matrix is ​​identical. This represents the total number of sampling time steps in the historical load data. The length of the sliding window;

[0100] Hankel embedding matrix for each load node Perform singular value decomposition to obtain the singular value diagonal matrix. :

[0101]

[0102] in, It is a left singular vector matrix. It is a singular value matrix. Let T be a right singular vector matrix, and let T be its transpose.

[0103] Preserving the singular value diagonal matrix after singular decomposition The former r The largest singular value and the corresponding left and right singular vectors , Construct a low-rank approximate embedding matrix:

[0104]

[0105] in, Indicates load node j The The low-rank approximate embedding matrix over the bar modal components, As a preset constant, They are respectively load nodes j The The first of the modal components p One singular value, a left singular vector, and a right singular vector;

[0106] And through the inverse Hankel operation Reconfigured as a load node j In the kLoad reconstruction signal on the i-th modal component: , obtain the multi-scale smooth reconstructed load feature tensor:

[0107]

[0108] wherein, denotes the load node j Load feature tensor on the i-th modal component at time t denotes the load node j Load reconstruction signal on the i-th modal component at time k t

[0109] Step S3 of the embodiment performs multi-scale time series structure modeling and noise reduction processing on the historical load data of multiple load nodes in the power system by combining multivariate empirical mode decomposition (MEMD) and singular value decomposition (SVD), aiming to extract more smooth and physically meaningful load feature tensors to provide higher quality input data for subsequent prediction models. Multivariate empirical mode decomposition (MEMD) is introduced to decompose the original load time series into several intrinsic mode function (IMF) components with different time scale characteristics. This process helps to capture the change characteristics of power load at different frequency levels (such as high-frequency disturbance and low-frequency trend), realizes multi-scale modeling of load signal, and solves the problem of difficulty in effectively extracting multi-time scale information from original load data. For each modal component, the dominant low-rank structure feature is obtained by Hankel matrix construction and singular value decomposition. Hankel embedding operation expands one-dimensional time series into a two-dimensional matrix with time delay structure, which helps to explicitly retain its time evolution law; and SVD can extract the main change direction from it, remove noise and local fluctuation information, and thus realize smooth reconstruction of the modal component. Through low-rank approximation, only the most representative dynamic mode in each modal is retained, improving the robustness and predictability of the data. The smooth reconstruction results of all modal components are spliced in the time dimension to form a complete multi-scale load feature tensor. This tensor not only retains the main trend and fluctuation mode of the original load signal, but also effectively filters out noise and abnormal disturbance, has higher learnability and stronger representation ability. This step realizes denoising, dimensionality reduction and multi-scale enhanced expression of historical load data, solves the problems of existing technologies such as violent fluctuation of load data, difficulty in handling abnormal disturbance, and mixing of different time scale changes, provides a solid data foundation for the model to accurately depict the load evolution law and improve the prediction performance under extreme weather, and significantly improves the accuracy and stability of the overall prediction.

[0110] ​​​Step S4: based on the physical connection relationship of each load node in the power system, a graph network structure and an adjacency matrix are constructed, and the load feature tensor and the meteorological feature vector of the corresponding node are spliced on each time slice to form the time sequence feature representation of the graph network node, specifically including:

[0111] Based on the meteorological feature vector and the load feature tensor of each load node, and the physical connection relationship, a graph network structure at a plurality of historical moments is constructed, specifically including:

[0112] All load nodes in the power system form a node set V V , and the physical connection relationship between the power transmission and distribution lines of each load node forms an edge set E E , wherein the node feature of the graph network structure corresponding to each moment is:

[0113]

[0114] Wherein, is the feature vector of the node of the graph network structure corresponding to moment t, is the load feature tensor of the load node j at moment t , and is the meteorological feature vector of the load node j at moment t , and represents the vector splicing operation.

[0115] The edge weight of the graph network structure corresponding to each moment is determined by the physical properties of the power transmission and distribution lines between the load nodes, and a normalized adjacency matrix with a self-loop is constructed :

[0116]

[0117]

[0118]

[0119] Wherein, represents the edge weight of node and , I is the unit matrix, N represents the number of load nodes, is the degree matrix, and the diagonal element ;

[0120] With the node set V, the normalized adjacency matrix , and the node feature , the graph network structure at each moment is defined obtaining a set of graph network structures .

[0121] Step S5: based on the constructed graph network structure and the adjacency matrix, constructing a power load prediction model based on a spatiotemporal graph convolution network, and realizing power load prediction based on the constructed power load prediction model;

[0122] The power load prediction model based on the spatiotemporal graph convolution network is as shown in Figure 2 , and includes:

[0123] The spatial graph convolution module is composed of multiple layers of stacked graph convolution layers (Graph Conv). Each graph convolution layer receives the node feature matrix and the normalized adjacency matrix of the time slice t as input, and outputs spatial features through graph convolution operation.

[0124] The time convolution module includes multiple layers of causal one-dimensional convolution (Causal Conv). Each layer of causal one-dimensional convolution convolves the node features of the same load node, and outputs time sequence features.

[0125] The spatiotemporal fusion module (Spatiotemporal Fusion) includes a feature splicing layer and a fusion mapping layer, which are used to fuse and map the spatial features output by the spatial graph convolution module and the time sequence features output by the time convolution module to the prediction dimension, to generate spatiotemporal features.

[0126] The output prediction submodule is used to map the spatiotemporal features of each load node after fusion to the load prediction value of the next time period through one or more fully connected layers (Fully Connected).

[0127] Based on the constructed graph network structure and the adjacency matrix, the power load prediction model based on the spatiotemporal graph convolution network is constructed, and specifically includes:

[0128] For each historical time t , the normalized adjacency matrix in the graph network structure and the node feature matrix are input into the spatial graph convolution module of the power load prediction model, graph convolution operation is performed, neighbor features are aggregated, and the spatial features of each node at time t are obtained:

[0129] The spatial features of each node from time to t are collected and input into the time convolution module to perform multiple layers of causal one-dimensional convolution or dilated convolution, capture short-term mutations and long-term trends between nodes, and output time sequence features of each node:

[0130] The spatial features and the time sequence features are spliced in the feature dimension, and are mapped to a unified dimension through a space-time fusion module to generate node-level space-time features;

[0131] The space-time features are input into an output prediction submodule to generate predicted load values of each node at the next time;

[0132] The actual load values of each node at the next time are used as labels to calculate the mean square error loss, and the parameters of the power load prediction model are iteratively updated according to the loss using back propagation and an optimizer to train the power load prediction model.

[0133] Embodiment 2:

[0134] The embodiment provides a power load prediction system based on a space-time graph convolutional network under extreme weather, comprising:

[0135] A data acquisition module is configured to acquire historical load data of multiple load nodes (i.e., substation nodes) in a power system and multi-dimensional meteorological element data of a corresponding area, wherein the meteorological elements include but are not limited to temperature, humidity, air pressure, wind speed, and precipitation.

[0136] A data preprocessing module is configured to clean, align, and normalize the acquired data, and to filter out meteorological features most closely related to power load through a combination of correlation analysis and model sensitivity analysis to improve the effectiveness and generalization ability of model input.

[0137] A feature extraction module is configured to perform multi-scale decomposition on load time series based on multi-variable empirical mode decomposition (MEMD), extract multi-order modal components, and perform singular value decomposition (SVD) and low-rank approximation reconstruction on each order of mode to obtain a smooth and denoised load feature tensor.

[0138] A graph construction module is configured to construct a graph structure adjacency matrix between power grid nodes based on geographical distance and load correlation, and obtain a graph representation available for graph convolutional networks through normalization processing.

[0139] A load prediction model module includes multiple parallel graph convolution (Graph Conv) + causal convolution (Causal Conv) units configured to extract spatial correlation features and time-dependent features between nodes, respectively. The spatio-temporal features of each branch are integrated through a fusion mechanism (Spatiotemporal Fusion), and the fused features are input into a fully connected network (Fully Connected) to output predicted load values.

[0140] A result output module is configured to output predicted load values of each node at a future time, and to compare the predicted load values with actual load values to further calculate error indicators for model evaluation and optimization.

[0141] If the above functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the present application that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of software products. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in various embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0142] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for power load forecasting based on spatio-temporal graph convolutional network under extreme weather, characterized in that, The method comprises the following characteristics: Collecting historical load data of a plurality of load nodes in a power system and regional meteorological element data; Based on the regional meteorological element data, the feature contribution value of each meteorological feature is calculated by combining model interpretation and regression analysis, and a plurality of key meteorological features are selected to construct a meteorological feature vector; The historical load data of a plurality of load nodes in the power system is subjected to multivariate empirical mode decomposition, a plurality of intrinsic mode function components are extracted, and singular value decomposition is performed on each mode component to construct a low-rank approximate embedding matrix and obtain a multi-scale smooth reconstructed load feature tensor; Based on the physical connection relationship of each load node in the power system, a graph network structure and an adjacency matrix are constructed, and the load feature tensor and the meteorological feature vector of the corresponding node at each time slice are spliced to form a time sequence feature representation of the graph network node; Based on the constructed graph network structure and adjacency matrix, a power load prediction model based on a spatio-temporal graph convolution network is constructed, and power load prediction is realized based on the constructed power load prediction model; The feature contribution value of each meteorological feature is calculated based on the regression coefficient and the model interpretation contribution value of each meteorological feature, and the feature contribution value of each meteorological feature to the load data is calculated: Based on the regional meteorological element data and the load data at the corresponding moment, a multiple linear regression model is constructed to obtain regression coefficients of each meteorological feature ; Based on the regional meteorological element data and the existing load prediction model, the model explanation contribution value of each meteorological feature is calculated by using a model explanation algorithm ; The load node is each transformer node, the historical load data is the active power value measured by each transformer node in a preset sampling period, the regional meteorological element data is a plurality of meteorological features measured in the same sampling period in the region where the transformer station is located, and the meteorological features include atmospheric temperature, dew point temperature, atmospheric pressure, relative humidity, wind speed, precipitation, and sunshine duration. wherein, is the feature contribution value of the i-th weather feature, i , is the regression coefficient of the i-th weather feature, and i is the model explanation contribution value of the i-th weather feature.​ For all meteorological characteristics Sort the values ​​from largest to smallest and select the first few. k Construct a meteorological feature vector based on the corresponding meteorological features. ; The multiple linear regression model is constructed based on the regional meteorological element data and the load data at the corresponding moment, and regression coefficients of each meteorological feature are obtained , and specifically comprises: Taking the load increment t at each sampling time as the dependent variable, and the meteorological feature vector at the same time as the independent variable, a multiple linear regression model is established: wherein, , are the load values at time t and time t-1 in historical load data, respectively, m is the total number of meteorological features, is the regression coefficient of the i meteorological feature, representing the degree of linear influence of the meteorological feature on the load increment; represents the sampling value of the i meteorological feature at time t , is an error term, is a constant term; The regression coefficients of each meteorological characteristic are obtained by least square method ; The model explanation algorithm is used to calculate model explanation contribution values of each meteorological feature based on the regional meteorological element data and the existing load prediction model , and specifically comprises: For the i-th weather feature, define a set of weather features as i For any subset of features i Divide the input vector into a weather feature vector containing the subset S and a weather feature vector containing the subset S and the feature to be tested i Calculate the model explanation contribution value of the i-th weather feature according to the existing load forecasting model i :​​​​​ wherein, denotes a subset S of the number of weather features, denotes a factorial of the number of weather features, denotes a factorial of the number of weather features, denotes a factorial of the total number of weather features, m , , denote an output of the existing load forecasting model when the input weather feature vector is and , respectively.

2. The method of claim 1, wherein the method is based on a spatiotemporal graph convolution network. The historical load data of a plurality of load nodes in the power system is subjected to multivariate empirical mode decomposition, a plurality of intrinsic mode function components are extracted, and singular value decomposition is performed on each mode component to construct a low-rank approximate embedding matrix and obtain a multi-scale smooth reconstructed load feature tensor, and the method comprises the following steps: 3.The power load forecasting method based on spatio-temporal graph convolution network in extreme weather according to claim 1, characterized in that, The power load prediction model based on the spatio-temporal graph convolution network comprises: In the power system N The historical load sequences of each load node are obtained by applying multivariate empirical mode decomposition. K A sequence of intrinsic mode functions ,in, Indicates load node j At any moment t The Bar modal components; each modal component for each load node with a sliding window length L constructing a Hankel embedding matrix: wherein, represents a load node j the first modal component of the matrix constructed after time-delay embedding; represents an operation that converts a one-dimensional time series into a two-dimensional matrix with a time structure, such that each anti-diagonal element in the matrix is the same, represents the total number of sampling time steps in the historical load data, is the length of the sliding window;​ Hankel embedding matrix for each load node performing singular value decomposition to obtain a singular value diagonal matrix : wherein is a left-singular vector matrix, is a singular value matrix, is a right-singular vector matrix, T is a transpose.

4. The method of claim 1 or 3, wherein, A time convolution module comprising a plurality of causal one-dimensional convolutions, each causal one-dimensional convolution being used for convolution on the node features of the same load node and outputting time sequence features; singular value diagonal matrix after singular value decomposition is reserved the first r largest singular values and the corresponding left and right singular vectors , , construct a low-rank approximation embedding matrix: wherein, represents a load node j the first rank-approximated embedding matrix on the first is a preset constant, are the first j singular value, left singular vector and right singular vector on the first singular component of the load node p , respectively. and the load reconstruction signal on the modal component is obtained by inverse Hankel operation reconstructing to the load node j In the first k the load reconstruction signal on the modal component: , obtaining the multi-scale smooth reconstructed load feature tensor wherein, denotes the load node j at time t on the load feature vector, denotes the load node j at the k th modal component at time t the load reconstruction signal.

5. The method of claim 1, wherein, A spatio-temporal fusion module comprising a feature splicing layer and a fusion mapping layer, which are used for fusing and mapping the spatial features output by the spatial graph convolution module and the time sequence features output by the time convolution module to a prediction dimension to generate spatio-temporal features; The spatial graph convolution module is composed of multiple layers of stacked graph convolution layers, each of which receives a node feature matrix and a normalized adjacency matrix of a time slice as input and outputs a spatial feature through graph convolution operation. t The spatial graph convolution module is composed of multiple layers of stacked graph convolution layers, each of which receives a node feature matrix and a normalized adjacency matrix of a time slice as input and outputs a spatial feature through graph convolution operation. An output prediction submodule used for mapping the fused spatio-temporal features of each load node to the load prediction value of the next time period through one or more fully connected layers. The graph network structure and the adjacency matrix are constructed based on the physical connection relationship of each load node in the power system, and the method comprises the following steps: Based on the meteorological feature vector and the load feature tensor of each load node and the physical connection relationship thereof, a graph network structure at a plurality of historical time points is constructed, and the method comprises the following steps:

6. The method of claim 1, wherein, Based on the constructed graph network structure and adjacency matrix, the power load prediction model based on the spatio-temporal graph convolution network is constructed, and the method comprises the following steps: ​ A node set is formed by all load nodes in the power system V A edge set is formed according to the physical connection relationship of power transmission and distribution lines between various load nodes E The node characteristics of the graph network structure corresponding to each moment are: wherein, the node of the graph network structure corresponding to time t the feature vector of the node, the load node j the load feature tensor at time t on the load node the meteorological feature vector of the load node j at time t on the load node denotes a vector concatenation operation; The edge weight of the graph network structure corresponding to each moment is determined by the physical properties of the power transmission and distribution lines between the passing nodes of the load nodes, and a normalized adjacency matrix with a self-loop is constructed : wherein, representing nodes with edge weights, is the identity matrix, N denotes the number of load nodes, is the degree matrix, whose diagonal elements ; With a node set V, a normalized adjacency matrix And node features , define the graph network structure at each time , obtain a set of graph network structures .

7. The method of claim 1 or 5 or 6, wherein, ​ For each historical moment t The normalized adjacency matrix In the graph network structure The node feature matrix The spatial graph convolution module of the power load prediction model is input, the graph convolution operation is performed, the neighbor features are aggregated, and the spatial features of each node at the moment t Are obtained: Collecting the spatial features of each node from time t, inputting a time convolution module, performing multi-layer causal one-dimensional convolution or dilated convolution, capturing short-term mutations and long-term trends between nodes, and outputting the time sequence features of each node: The spatial features and the time sequence features are spliced in a feature dimension, and are mapped to a unified dimension through a space-time fusion module to generate node-level space-time features; The space-time features are input into an output prediction submodule to generate predicted values of the loads of the nodes at the next moment; The mean square error loss is calculated by taking the real load values of the nodes at the next moment as labels, and the parameters of the power load prediction model are iteratively updated according to the loss by using back propagation and an optimizer, so that the power load prediction model is trained.

Images