A short-term residential load forecasting scheme based on the multiple correlation-temporal graph neural networks

Abstract

Accurate residential load forecasting (RLF) is of great significance for the decision-making and operation of modern power system. In literature, deep neural network (DNN) based RLF schemes have witnessed great development due to the advantage of automatically extracting features and capturing complex non-linear pattern in the presented data. However, most existed works separately exploit the historical data of a specific residential house to forecast its load. However, the electricity consumption behaviors among residential users are not independent, and implicitly have some correlations, which can be explicitly characterized and exploited to improve the accuracy of RLF forecasting. Inspired by this idea, through exploiting the multiple correlations among households, this paper proposes a novel residential load forecasting framework based on multiple correlation-temporal graph neural networks, RLF-MGNN. Specifically, the novelty of our work includes three aspects. First, multiple graphs are explicitly constructed to represent both linear and nonlinear correlations among temporal load series of households. That is, the synchronization graph is built to describe the degree of linear correlation between two households using Pearson correlation coefficient, which characterizes the similarity of their consumption behaviors, and the causality graph is built to describe their nonlinear correlation using transfer entropy, which characterizes the amount of directional information transfer from one time series to another, and models the mutual influence between households. Second, the multiple correlation-temporal graph convolutional networks (GCNs) are designed to forecast the residential users’ loads. In detail, at each timestep, latent features are first extracted by corresponding GCNs to embed multiple correlations among households, and then are sent to Long Short-Term Memory (LSTM) for further learning the latent temporal features. Finally, thorough experiments on real datasets demonstrate that our proposed RLF-MGNN outperforms the state-of-the-art independent DNN based schemes and other GNN based schemes.