Graph attention enhanced transformer network: an investigation into day-ahead electricity demand forecasting

Abstract

The evolution of modern power grids, also known as smart grids, along with distributed renewable power generation and demand-side management strategies, has increased variability in both electricity supply and consumption patterns. As power generation systems transition to smart grids, accurate day-ahead load forecasting has become critical for ensuring grid stability, operational efficiency, and cost-effective energy management under growing electricity demand. Traditional load forecasting approaches often struggle to model the nonlinear and temporal dependencies arising from dynamic weather conditions, shifting consumer behaviour, and increasingly diverse generation portfolios. This thesis proposes a novel approach based on the Temporal Fusion Transformer (TFT) architecture integrated with a Gated Attention Network (GAT) to analyze and predict electricity demand in Ontario. TFT models effectively identify the importance of individual features, but they fail to capture the direct interaction between pairs of features. The integration of a GAT allows the model to overcome this limitation by learning cross-feature relationships adaptively. The model leverages attention mechanisms, interpretable temporal dynamics and inter-feature representations to improve forecasting accuracy and provide insights into key influencing factors. The proposed model, trained on the Independent Electricity System Operator (IESO) dataset, demonstrates its effectiveness in handling the complexities of modern energy systems by achieving an MAPE of 2.92% and 2.18% on Ontario and market demand forecasting, respectively. Further evaluations on the ENTSOE, AEP, and ELIA datasets demonstrate state-of-the-art (SOTA) or near- SOTA performance, achieving an MAPE of 1.77%, 3.05%, and 2.45%, respectively, while validating model generalizability. The contribution of this thesis highlights the importance of capturing inter-feature and temporal dependencies in the field of day-ahead forecasting for smart grids.