Transfer learning and explainable AI for heating load forecasting: A large-scale benchmark with SHAP-based static features

Accurate short-term heating load forecasting is crucial to achieving flexibility in an increasingly electrified heating sector. However, the lack of data on individual buildings poses a major challenge. This study proposes a transfer learning (TL) framework that leverages SHAP-based feature importances and load-pattern metrics as static features (also called static covariates), combined with advanced TL strategies, to improve performance under limited data conditions. Across 18 mostly real heating load datasets and seven time series models—MLP, LSTM, TFT, PatchTST, TiDE, TimeMixer, and Time-LLM—fine-tuning with freezing (FTF) reduces median NRMSE by 22 % compared to the strategy no transfer learning (NTL), with improvements reaching up to 46.6 % for specific models such as TiDE. This benefit persists even with only 10 % of the training data. Another strategy zero-shot (ZS) approach—applying the pre-trained models without training on the target data—achieves a comparable reduction of 20 % on average. PatchTST and TiDE are particularly efficient at fine-tuning, combining high accuracy with low computational cost. Static features derived from SHAP-based feature importances and load-pattern metrics yield only minor, often statistically non-significant, and sometimes slightly negative effects for most architectures. This indicates that temporal dynamics already capture the dominant transferable patterns, and that explicit feature importance metrics offer limited additional predictive value in well-pre-trained TL settings. These findings confirm that well-designed TL strategies—especially FTF and ZS—can reliably compensate for limited data in heating load forecasting. Future work should explore adaptive freezing schemes, learning rate adaptation, and the transferability of the approach to longer forecast horizons. • Transfer learning (TL) reduces median NRMSE by 22 %. • PatchTST and TiDE remain robust with only 10 % training data. • PatchTST achieves good accuracy with minimal memory requirements. • Static features show no consistent accuracy gains across settings. • TL strategy is a major driver of accuracy in this benchmark.