Graph convolutional network-based aggregated demand response baseline load estimation

Abstract

Accurate aggregated baseline load (ABL) estimation is very important for the demand response (DR) compensation settlement between system operators and DR aggregators. Current ABL estimation methods totally ignore the spatial correlation between load patterns of different customers, which will lead to large errors when the load pattern on the DR event day fluctuates largely compared with that in historical days. To this end, this paper proposes a Graph Convolutional Network (GCN)-based ABL estimation method to improve the estimation accuracy. The basic idea is to enhance the estimator's ability to capture load uncertainty by sharing load fluctuation information between different customers. The proposed method contains three main steps: First, all customers are grouped into different clusters by the K-means algorithm according to their historical typical load patterns (TLPs). Second, these clusters are transformed into an undirected graph based on an adjacency matrix reflecting the spatial correlations, which are constructed according to the difference between the TLPs in different clusters. Third, the ABL estimation is transformed into a node regression problem of the graph. Case studies on a real load dataset verify the effectiveness and superiority of the proposed method.