Fair Benchmarking in Short‐Term Load Forecasting

Abstract

Performance comparisons in short-term load forecasting are often confounded by differences in preprocessing pipelines rather than reflecting intrinsic architectural capability. Variations in feature engineering, scaling, temporal windowing and data partitioning can dominate reported accuracy and obscure the actual behaviour of forecasting models. This study examines preprocessing–architecture interaction by benchmarking random forest, LightGBM, long short-term memory (LSTM), transformer and Temporal Fusion Transformer (TFT) under a shared tabular preprocessing pipeline, ensuring strict control over data handling and evaluation conditions. Under this controlled setting, tree-based models exhibit strong predictive performance, whereas deep sequence models experience substantial degradation when temporal continuity is not explicitly represented. To isolate architectural sensitivity from preprocessing effects, we further conduct a within-architecture analysis by retraining an identical LSTM under a sequence-aware pipeline aligned with its temporal inductive bias. This realignment yields an order-of-magnitude reduction in RMSE, demonstrating that preprocessing design is a first-order determinant of deep sequence model performance. The results establish a transparent and reproducible benchmarking framework and highlight the importance of aligning data representation with model assumptions when interpreting comparative performance in time series forecasting.