The Decomposition Penalty: Quantifying the Impact of Classical Pre-processing on Time-Series Foundation Models in Load Forecasting

The integration of Time-Series Foundation Modelsinto short-term load forecasting presents new opportunities formodern smart grid management. However, traditional power gridforecasting heavily relies on Seasonal-Trend (STL) decompositionas a mathematical pre-processing step. This study investigatesthe empirical impact of classical decomposition on zero-shotfoundation models (Chronos-T5) compared to traditional autoregressive gradient boosting (XGBoost). Testing across 107 rollingwindows of hourly power load data, we identify a severe ”decomposition penalty.” Stripping seasonality and trend degraded thefoundation model’s Mean Absolute Percentage Error (MAPE)from 9.25% to 13.15%. Conversely, a computationally efficient,purely autoregressive XGBoost model achieved a superior 7.68%MAPE. The empirical results suggest that classical decompositiondestroys the macro-structural context that foundation modelsrely on, and that tree-based ensembles currently remain a highlyrobust, computationally efficient architecture for load forecasting.