Graph Convolutional Networks based short-term load forecasting: Leveraging spatial information for improved accuracy

Accurate short-term load forecasting is vital for the efficient operation of the power sector. The challenge of predicting fine-resolution load, such as weekly load, is compounded by its inherent volatility and stochastic nature. However, forecasting becomes more tractable at higher scales, such as user clusters, where fluctuations are smoothed out. Existing methods focus solely on temporal data and auto-regressive processes for load prediction, disregarding the spatial information inherent in the power grid's graphical structure. This research proposes an innovative approach that integrates spatial and temporal information for short-term load forecasting. A novel technique is introduced to convert load data into a graphical representation, which is then processed by Graph Convolutional Networks (GCN) to capture spatial embeddings. These GCN embeddings, in conjunction with temporal features, are employed for load prediction. Rigorous experimentation employing advanced machine learning and deep learning techniques validates the effectiveness of the proposed approach. The findings reveal that leveraging spatial information through GCN embeddings significantly enhances load forecasting performance, leading to improvements of up to 39% which emphasize the potential of proposed method.