Graph Spatio-Temporal Attention Network-based Electricity Demand Forecasting

Abstract

Electricity demand forecasting is essential for improving the efficiency of power systems. Nevertheless, multi-step electricity demand forecasting is highly challenging due to the high volatility and uncertainty involved. In this paper, we investigate the spatio-temporal characteristics of electricity load and propose a Graph Spatio-Temporal Attention Network (GSTAN) to forecast multi-step electricity consumption of different users. We use the self-attention mechanism in temporal and spatial dimensions simultaneously, so that GSTAN can not only capture the temporal correlations but also the spatial correlations. Specifically, GSTAN adopts an encoder-decoder architecture, consisting of multiple spatio-temporal attention blocks to model the impact of the spatio-temporal correlations. First, we construct a spatial relationship graph based on the similarity of users' patterns to model spatial correlations. Second, we encode the spatio-temporal characteristics of users by using the decoder. Then the features generated by the encoder are transformed into the decoder by the transform attention layer, and the prediction sequence is finally output by the decoder. Experimental results on the realworld power consumption dataset demonstrate that our model performs better than state-of-art algorithms.