Transfer Learning for Residential Electricity Load Forecasting in a Multi-Geographical Context

Accurate electricity load forecasting is essential for grid management and energy market optimization. Traditional models require region-specific training, making them costly and data-intensive. This study explores transfer learning for residential electricity load forecasting, applying a Gated Recurrent Unit (GRU)-based model to three Danish cities-Fredericia, Strib, and Nr. Bjert-differing in dataset size, household composition, and seasonal patterns. Two transfer learning strategies are evaluated: direct application of a pre-trained model and partial fine-tuning on the target dataset. Results show that training on multi-year datasets significantly improves generalization, while models trained on short-term data struggle to adapt. These findings highlight transfer learning's potential to reduce model development costs and improve scalability in regions with limited historical data. The study provides insights into successful transfer learning conditions, with direct implications for energy utilities and policymakers. It demonstrates how transfer learning enhances forecasting accuracy and streamlines sector-coupled energy planning. Future research should explore alternative deep learning architectures, exogenous feature integration, and cross-regional applications.