Spatio-Temporal Short Term Load Forecasting Using Graph Neural Networks

Accurate short-term load forecasting (STLF) is essential for the efficient operation of the power sector. Due to heightened volatility and intrinsic stochasticity, forecasting load at a fine resolution, such as weekly load, is difficult. Existing STLF techniques only rely on temporal data and auto-regressive processes to forecast load. However, the power grid has a graphical structure that provides spatial information too. This paper proposes an innovative STLF method fusing both spatial and temporal information. We propose a creative way to convert load data into graphical form, which is fed into graph convolutional networks (GCN) to learn spatial embeddings. The GCN embeddings are used along with temporal features to predict the load. We perform extensive experiments using state-of-the-art machine learning and deep learning techniques to validate our approach. The results demonstrate that by using spatial information, we can sub-stantially improve the forecasting performance.