Instance-Based Transfer Learning for Short-Term Load Forecasting in Data-Scarce Buildings

Accurate short-term load forecasting is essential for energy-efficient building operations, however, many buildings lack sufficient historical data to train reliable predictive models. This paper investigates the use of instance-based transfer learning to improve forecasting performance in data-scarce buildings by reusing models trained on similar, data-rich buildings. We cluster buildings of various types based on their energy consumption profiles using dynamic time warping and hierarchical clustering. Within each cluster, we fine-tune XGBoost and LSTM models pretrained on source buildings (with sufficient data) to forecast the consumption in target buildings (with limited data) with access to increasing percentage of their original limited data of 7, 14, 21, 28, ... , 245 days. Our evaluation compares the performance against models trained from scratch using the same limited data. Our results show that transfer learning significantly improves forecasting accuracy in highly data-constrained settings, reducing RMSE by up to 40%. However, as more training data becomes available, its benefits diminish, and baseline models eventually outperform transferred ones beyond an average threshold of 140–168 days.