Original-scale calibrated probabilistic building electricity load forecasting for demand-side management

Abstract

Reliable day-ahead electricity forecasts are needed for building energy management, demand-side operation, and risk-aware scheduling. In common data-driven pipelines, targets are normalised for training, whereas forecast errors and uncertainty intervals are interpreted after inverse transformation to kWh. This creates a cross-space mismatch for heterogeneous building loads. We propose DSRC-TFT, a lightweight calibration extension of the Temporal Fusion Transformer that combines original-scale interval regularisation, linear residual refinement, and horizon-aware quantile calibration while retaining the TFT backbone. On BDG2, DSRC-TFT reduces MAE, RMSE, WAPE, pinball score, Winkler score, and prediction-interval width relative to TFT and quantile-regression baselines, with empirical coverage remaining close to the nominal 90% level. Sensitivity analysis reveals an accuracy–sharpness–coverage trade-off rather than a universally optimal interval-loss weight. Five independent random initialisations show limited variation in the main test metrics. Computational profiling further shows that the added modules increase the parameter count by less than 0.4% and profiler-reported FLOPs and inference latency by about 0.1%. The results support original-scale calibration as a practical way to improve probabilistic building electricity forecasts without materially increasing model complexity.