Multi-mode power demand prediction method based on space-time dependent learning

Abstract

The invention relates to a technology in the field of neural network application, in particular to a demand prediction method based on multimode power adaptive space-time dependent learning, which comprises the following steps of: constructing a power decision network in an offline stage and randomly initializing a modeled demand prediction model; the demand prediction model is composed of a feature representation module, a time network module, a space network module, a space-time dependence network module, a space-time fusion network module, a demand prediction network module and a loss calculation module. Performing feature representation on the multimode power flow data by using a feature representation module, forming a core of a demand model through a time network module, a space network module, a space-time dependence network module and a space-time fusion module, and forming a demand prediction model; and in the online use stage, through a demand prediction model obtained through training, power flow demand prediction is realized based on the input multimode power state data.

Description

Technical Field

[0001] The present invention relates to a technology in the field of neural network applications, specifically a demand prediction method based on multi-mode power adaptive spatio-temporal dependence learning. Background Art

[0002] Most of the existing demand prediction technologies focus on improving the accuracy of power demand prediction in a single spatial relationship of a single power mode, ignoring the diversity and heterogeneity of the spatial relationships of power stations in the power network. Its technical problems include: difficulty in fusing multiple power modes for power demand prediction; difficulty in capturing spatial connections from multiple perspectives; and difficulty in modeling dynamic time dependence. Summary of the Invention

[0003] In view of the above deficiencies of the existing technology, the present invention proposes a multi-mode power demand prediction method based on spatio-temporal dependence learning.

[0004] The present invention is realized through the following technical solutions:

[0005] The present invention relates to a multi-mode power demand prediction method based on spatio-temporal dependence learning. In the offline stage, a power decision network is constructed and the modeled demand prediction model is randomly initialized. The demand prediction model consists of a feature representation module, a time network module, a spatial network module, a spatio-temporal dependence network module, a spatio-temporal fusion network module, a demand prediction network module, and a loss calculation module. The feature representation module is used to perform feature representation on multi-mode power flow data. Through the time network module, the spatial network module, the spatio-temporal dependence network module, and the spatio-temporal fusion module, the core of the demand model is formed to form the demand prediction model. In the online usage stage, based on the input multi-mode power state data, the power flow demand prediction is realized through the trained demand prediction model.

[0006] The present invention relates to a demand prediction system based on multi-mode power adaptive spatio-temporal dependence learning for implementing the above method, including: a feature representation module, a time network module, a spatial network module, a spatio-temporal dependence network module, a spatio-temporal fusion network module, and a loss calculation module. Among them: the feature representation module performs feature representation processing according to different mode power state data (such as weather, current power grid state, power congestion degree, etc.) in the training data, and outputs the feature representation of different mode power flow data; the time network module realizes time alignment for the input multi-mode power state data; the spatial network module realizes spatial alignment for the input multi-mode power state data; the spatio-temporal dependence network module realizes spatio-temporal aggregation for the same mode power state data that has been time and space aligned; the spatio-temporal fusion network realizes overall fusion for the aggregated power state data output by the spatio-temporal dependence network of all modes; the loss calculation module realizes the calculation of the two-norm loss between the power flow prediction result and the real power flow data, serving for model training.

[0007] The described feature representation module includes: a data preprocessing unit, a multi-source fusion unit, a time feature extraction unit, and a spatial embedding unit, where: The data preprocessing unit fills in missing values, eliminates outliers, and performs normalization processing on the input original multi-modal power status data (such as weather, grid load, regional energy consumption, etc.) to obtain a regularized time series data matrix; The multi-source fusion unit, based on the preprocessed heterogeneous data (such as meteorological data, economic indicators, historical demand sequences), adopts a feature-level fusion strategy, maps features of different modalities to a unified latent space through a fully connected layer, and obtains a fused feature tensor; The time feature extraction unit captures local time patterns using a one-dimensional convolutional network based on the fused feature tensor, and introduces a long short-term memory network to model long-term time series dependencies, and outputs time-enhanced features; The spatial embedding unit generates site-level spatial vectors using graph embedding technology based on the grid topology structure data, and forms a spatio-temporal joint feature representation after splicing with the time-enhanced features.

[0008] The described spatial learning module includes: a graph construction unit, a neighborhood aggregation unit, a multi-head attention unit, and a spatial feature update unit, where: The graph construction unit constructs a dynamic adjacency matrix based on the virtual site clustering results and the grid physical connection relationship, and defines the spatial dependence weights between sites; The neighborhood aggregation unit performs weighted aggregation on the neighborhood features of each site based on the graph attention network; The multi-head attention unit uses K groups of independent attention mechanisms to perform parallel calculations, and splices the outputs of each group and then performs a linear transformation to enhance the robustness of the spatial features; The spatial feature update unit fuses the original features and the aggregated features through residual connection and layer normalization, and outputs a high-order spatial dependence representation.

[0009] The described spatio-temporal attention module includes: a time series slicing unit, a multi-head self-attention unit, a period fusion unit, and a spatio-temporal gating unit, where: The time series slicing unit divides the input sequence into long-term, medium-term, and short-term segments, and extracts dependence patterns of different time granularities respectively; The multi-head self-attention unit performs dot product attention calculations on each time series segment; The period fusion unit splices the attention outputs of different time granularities along the channel dimension, and realizes cross-period feature interaction through a convolutional layer; The spatio-temporal gating unit dynamically adjusts the contribution weights of spatio-temporal features using a gated linear unit.

[0010] The described encoder includes: a spatial Transformer module, a time Transformer module, where: The spatial attention module extracts the dynamic spatial dependence between stations within a certain period of time. The time attention module models the time dependence of each independent station.

[0011] The results obtained by the described spatial Transformer module and time Transformer module are fused through a gating mechanism. The specific process is as follows: 1) Feature concatenation: Concatenate the dynamic spatial dependency feature matrix output by the spatial Transformer and the temporal dependency feature matrix output by the temporal Transformer along the feature dimension to obtain a joint feature matrix; 2) Gating weight generation: Calculate the dynamic fusion weights of spatial and temporal features through a learnable gating weight matrix and bias terms; 3) Feature weighted fusion: Perform weighted summation on spatial features and temporal features; 4) Residual enhancement: Add the original input features and the fused features, and stabilize the training process through layer normalization.

[0012] The decoder described above includes: a temporal embedding unit, an attention weighting unit, a linear prediction unit, and a multi-task fusion unit, where: the temporal embedding unit combines the historical demand sequence with the position encoding of future timestamps to generate decoder input embeddings; the attention weighting unit calculates the similarity weights between historical time series and predicted time series through a cross-attention mechanism; the linear prediction unit performs an affine transformation on the attention-weighted features and outputs the demand prediction values for each station; the multi-task fusion unit integrates the prediction results of different power modes (such as wind power, thermal power, and user-side demand) and generates the final demand distribution through weighted summation. Detailed implementation 1) This embodiment relates to a demand prediction method based on multi-mode power adaptive spatio-temporal dependency learning. According to the power network in a specific area and the additional features of historical departure and arrival demands of multiple power modes, the goal of power demand prediction is to predict the future departure and arrival demands of the corresponding power modes in the next time interval. The goal is to construct a mapping function h(·): where represents the estimated power demand value at the next time step. The power demand of a single power mode can be defined as the quantity entering and leaving the area within a certain time interval. p , , i , i , (t) , i , (t) , i , , (t) , i , , (t) , (t) (t) and d i (t) respectively record the historical observed values of the boarding and alighting demands at the station at the time step. To study the correlation between boarding and alighting behaviors, these two historical time series are concatenated simultaneously and denoted as X i (t) = [p i <​​​​​​​​3) Step 1: Divide the city into an m×n grid, and construct sets X and Y to record the pick-up and drop-off demand values of each grid and the distance of each grid, specifically: Divide the city into an I×J grid, and construct two sets R and U, R = {ρ1, ρ2, …, ρ I×J}, U = {δ1, δ2, …, δ I×J}, to record the pick-up and drop-off demand values of each grid and the distance of each grid 4) Step 2: Use the density peak clustering method to construct virtual stations and divide all the grids obtained in Step 1 into the corresponding clusters, specifically including: 5) 2.1) Regard the cluster centers constructed in the algorithm as virtual stations, and regard the number of clusters as the number of virtual stations, where the cluster centers satisfy: the local density of the cluster center is greater than the density of adjacent clusters and the distances between different cluster centers are far. 6) 2.2) Construct virtual stations by using the density peak clustering-based algorithm. According to the theory that the cluster to which a grid belongs is the same as the cluster to which its nearest neighbor belongs, divide all the grids into the corresponding clusters. 7) 2.3) Learn the power conversion between stations through the moving graph, which helps to capture the correlation between the pick-up and drop-off power demands, specifically: Among them: represents the trip record from station i to station j within a period of time. 8) 2.4) Study the correlation between stations when the power demand changes over time through the similarity graph to form a distance graph Specifically: represents the passenger flow sequence of the i-th station in the t-th month. 9) Step 3: Use the graph attention mechanism GAT to construct a spatial learning module and calculate the weight moment matrix, specifically including: 10) 3.1) Calculate the relationship R between the nodes in the input H (t) ∈R N×F of the graph, where R ij represents the relationship between node i and node j, which is a learnable parameter, and 11) 3.2) Based on the relationship R and the original power graph, calculate the weight matrix A, and use the adjacency matrix to further process the relationship R, specifically: 12) 3.3) To make the learning process more stable, adopt the multi-head attention mechanism, use K independent attention mechanisms, and average the results to obtain 13) 3.4) The constructed spatial learning module, according to the input X ∈ R N×F , where N is the number of nodes, F is the number of input features, and the output is X G ∈ R N×F' , obtaining X G = GATT(X, G), that is, respectively based on the graph attention mechanism GAT, the input X is executed Then concatenate them to obtain 14) Step four: Construct a spatio-temporal attention module, and use a standard multi-head self-attention mechanism and a position fully-connected feed-forward network to process the output of the spatial learning module in step three to obtain a spatio-temporal fusion data representation. Specifically: Use a standard multi-head self-attention mechanism and a position fully-connected feed-forward network. In addition, different time periods are provided in the attention mechanism, which helps to associate the previous long-term and short-term time series. The output of the attention mechanism, the input of the attention mechanism usually contains a query Q and a set of key-value pairs K, V, and then the output is calculated as a weighted sum of these values. Considering the advantages of dot-product attention: faster training speed and more space-saving, use the dot product and multiply the dot product by to counteract the extremely small gradient phenomenon. Compared with single-head attention, it is effective to use different learning queries, keys, and values to perform the multi-head attention function. MultiHead(Q, K, V) = [h1, h2, …, h n W o , h i = Attention(QW i Q , KW i K , VW i V ), in addition, applying the fully-connected feed-forward network (FFN) to each position is the same, which can capture the dependencies of nearby sites at the same time. This network consists of two linear transformations with a ReLU activation function in the middle: FFN(x) = max(0, xW1 + b1)W2 + b2. Considering that the long-term attention mechanism cannot effectively capture short-term dependencies, select three periods using the attention mechanism, and then concatenate these results. Through experiments with different numbers of cycles, select β = 3, T = {3, 6, 12}, so as to obtain better prediction results: Then, based on the fusion of the output of the attention mechanism, the value of the embedding module, and the value of the residual connection, the output of the time attention mechanism module is obtained: 15) Step Five: Perform fusion processing on the output of the attention mechanism in Step Four to obtain the output of the temporal attention mechanism module constructed in Step Five, specifically as follows: The output result of the entire spatial attention mechanism Where: 16) Step Six: Construct a decoder. By performing a linear transformation on the output result of the spatial attention mechanism obtained in Step Five, further predict a specific requirement obtained in Step Five to obtain a prediction result, specifically as follows: Based on the embedding unit in the decoder, embed the historical time series and the future time series, and then apply the attention mechanism to calculate the importance score between the historical time series. And define the prediction time step t j And the historical time step t i ∈t similarity: Then perform normalization through softmax: Further perform weighted combination on the normalized attention coefficients to obtain the representation form of the prediction time step t j . Through linear transformation, predict a specific requirement Y′ = W O F O + b O . And through the L2 loss function, set Control the weight to obtain: 17) Step Seven: Set the control weight to obtain the final prediction result. 18) The above specific implementation can be locally adjusted by those skilled in the art in different ways without departing from the principles and purposes of the present invention. The protection scope of the present invention is subject to the claims and is not limited by the above specific implementation. All implementation solutions within its scope are subject to the present invention.

Claims

1. A demand forecasting method based on multi-mode power adaptive spatio-temporal dependence learning, characterized in that Construct the power network in the offline stage and randomly initialize the multi-mode power decision network. Use the adaptive spatio-temporal dependence learning algorithm to group different power modes, form local communication groups and train the decision network. In the online stage, through the trained decision network, dynamically guide the power modes to perform grouped communication in the demand forecasting task in real time, decompose the overall demand forecasting task into the aggregation of several local tasks, and achieve demand forecasting and optimization.

Images