Stacking a BiLSTM-Attention Network With a Pretrained Time-Series Foundation Model and Gradient-Boosted Residual Correction for Short-Term Electricity Load Forecasting: A Comparative Evaluation on the Tetouan City Dataset
This study investigates whether adding a pretrained time-series foundation model (Chronos-Bolt) as a second forecaster alongside a domain-specific Bidirectional Long Short-Term Memory (BiLSTM)-attention network, combined through a learned gradient-boosted correction layer, improves short-term electricity load forecasting compared with a recently published hybrid LSTM-XGBoost baseline using the Tetouan City dataset. A BiLSTM with additive attention was trained on 24-hour rolling windows of weather, calendar, and load-lag features for Zone 2 of the Tetouan City dataset (52,416 ten-minute readings from 2017). Chronos-Bolt-Small was applied in a zero-shot setting, forecasting from raw load history alone without fine-tuning. A Light Gradient-Boosting Machine model combined both forecasts with the engineered features to learn a residual correction on top of the stronger anchor forecast, with early stopping based on a held-out validation slice. Two early design choices reduced accuracy before the final version was reached: applying principal component analysis to the exogenous features and training the meta-model to predict load directly rather than a correction. Both issues were diagnosed and addressed. Separately, a validation-checkpointing problem that distorted the deep-learning ablation ranking was corrected by evaluating the models six times per training epoch instead of once. The final pipeline achieved a root mean square error (RMSE) of 391.68 kWh, a mean absolute percentage error (MAPE) of 1.21%, and an R² of 0.9949 on a chronological 80/20 Zone 2 split. This represents a 43% lower RMSE and approximately 50% lower MAPE than the figures reported for the published baseline (which was not re-run here; see Limitations), and an 18% improvement over Chronos-Bolt used alone. A Diebold-Mariano test against an internally trained plain LSTM indicated a statistically significant difference (Diebold-Mariano test = 67.07, p < 10⁻⁶). After the validation fix, the ablation study showed improvements in RMSE and R² at every stage, with MAPE following the same overall trend apart from a small increase at the BiLSTM step. When run without the foundation-model signal, the pipeline did not consistently match the published Zone 2 figures (Zone 2 without Chronos: 955.93 kWh; Zone 3: 678.06 kWh; Zone 1: 1,433.77 kWh), indicating that most of the improvement over the published baseline comes from the foundation-model forecast. SHapley Additive exPlanations showed that the Chronos forecast was the dominant input to the correction layer, with short-term load volatility and time-of-day features contributing smaller but consistent signals. The combination of a pretrained, covariate-free foundation model and a domain-specific deep network through a residual correction layer achieved lower errors than those reported for a published hybrid baseline on this dataset, and the foundation model proved to be the strongest individual forecaster tested. The comparison with the external baseline is indicative rather than a fully controlled experiment, and the results are limited to a single dataset and forecast horizon.
Authors
- Khushvir Singh (ORCID: https://orcid.org/0009-0006-4677-8047)
- Anurag Tewari (ORCID: https://orcid.org/0000-0002-1456-649X)
Institutions
- Lovely Professional University (IN)
Publication Details
- Journal
- Cureus Journal of Computer Science.
- Published
- 2026-11-20
- DOI
- https://doi.org/10.7759/s44389-026-00383-6
- Primary Topic
- Energy Load and Power Forecasting
- Type
- article
- Field-Weighted Citation Impact
- 0.00