Short-term spatial-temporal energy forecasting in a Danish distribution grid using a hybrid transformer-graph neural network model

The integration of renewable energy resources and the increasing need for grid balancing have intensified the demand for accurate short-term consumption and production forecasting in electricity grids. This progress has been facilitated by greater data availability and the development of advanced machine and deep learning algorithms. While models such as the Transformer have revolutionised temporal sequential data processing, spatial dependencies between nodes in electrical grids could improve predictions. This study proposes a spatial-temporal hybrid model that integrates a Graph Neural Network with a Transformer to capture both temporal and spatial relationships. The results show that the Transformer outperforms benchmark models, Long Short-Term Memory and Multilayer Perceptron, for consumption forecasting. Furthermore, incorporating a Graph Neural Network enhances consumption prediction accuracy. For production forecasting, the standalone Transformer slightly outperforms the hybrid model, suggesting less useful spatial correlations among production nodes. However, the hybrid model reduced trainable parameters by 57 %, making it more compact with minimal accuracy loss. • A hybrid Transformer-GNN model is proposed for energy forecasting. • The model captures both temporal and spatial dependencies in the grid. • GIN improves consumption forecasting via injective aggregation. • The hybrid model outperforms LSTM and MLP in consumption forecasting. • The hybrid model reduces trainable parameters by 57% with minimal accuracy reduction.