A Patch-Based Transformer for Building Energy Load Forecasting

A considerable portion of worldwide energy usage and carbon emissions comes from buildings. Accurate short-term electricity load forecasting is critical to optimize building energy management systems. In this study, the PatchTST model was evaluated for short-term electricity load forecasting at the building scale. This forecasting problem was addressed using 50 buildings from the public Building Data Genome Project 2 dataset. Experiments were conducted using a strict temporal split strategy, where models generated 24-hour forward predictions from 336 hours of historical data. Temporal features, including hour, day, month, and weekend indicators, were integrated into the PatchTST model. The model was compared against Persistence, Long Short-Term Memory (LSTM), and DLinear baselines. On the test set, the PatchTST model outperformed all baselines, achieving a mean absolute error (MAE) of 7.24 kWh, a root mean square error (RMSE) of 13.85 kWh, an R² of 0.9894, and a symmetric mean absolute percentage error (MAPE) of 9.42%. Per-horizon analyses showed that PatchTST maintained stable performance across the forecast horizon. A building-level Wilcoxon signed-rank test indicated that the superiority of PatchTST over all baselines was statistically significant (p < 0.001) with large effect sizes These findings suggest that patch-based Transformer architectures provide high accuracy, statistical reliability for short-term electricity load forecasting.