Graph Convolutional Network Based Day-ahead Demand Response Potential Forecasting Model Considering the Spatial-temporal Correlation

Abstract

Accurate forecasting of demand response (DR) potential is of great significance for load aggregators (LAs)to bid in the DR market and reduce their market trading risks. Current DR potential forecasting methods usually ignore the spatial correlation between different types of customers, resulting in large errors when the DR potential varies dramatically. To this end, this paper proposes a day-ahead DR potential forecasting model for LAs based on a graph convolutional network (GCN), considering the spatial-temporal correlation between different types of customers. Firstly, customers are divided into different clusters using the K-means algorithm. Secondly, the multiple influencing features of each customer cluster are extracted and the spatial-temporal correlation matrices are established by Pearson correlation coefficients (PCCs). Finally, this graph structure with spatial-temporal correlation information is used to train the GCN forecasting model. Case studies verified the validity of the proposed DR potential forecasting model.