Time series models improve electrical load forecasting with fewer tweaks
Parameter-Efficient Adaptation of Time Series Foundation Models for Electrical Load Forecasting
Machine LearningArtificial Intelligence
Summary
Forecasting how much electricity people will use helps manage power systems better, but making new models for each situation can be expensive. The authors studied large pretrained time series models that can predict electricity use across different areas and building types. They found that updating only a small part of these big models (using a method called LoRA) can nearly match full retraining accuracy, saving time and computing work. This approach works well especially for forecasts in some local areas and buildings and balances accuracy and efficiency.
What this means in practice
- •For energy utility teams: Use compact adapted pretrained models to improve regional and building-level electricity demand predictions efficiently.
- •For smart building operators: Deploy parameter-efficient adapted models to forecast local electricity consumption with better accuracy while minimizing computational costs.
Authors
Aliullah, Julian Oelhaf, Edouard Fouché, Jessica Deuschel, Christian Bergler, Andreas Maier, Siming Bayer
Abstract
Accurate electrical load forecasting supports power-system operation across countries, regions, cities, and buildings, but repeated development of target-specific models is costly. Time series foundation models (TSFMs) provide reusable pretrained backbones, although multi-domain pretraining may not fully capture load-specific characteristics. This paper compares pretrained inference, full fine-tuning, and low-rank adaptation (LoRA) in terms of forecasting accuracy and adaptation efficiency across four electrical load aggregation levels. MOIRAI 1.0 Large serves as the primary backbone, while Chronos-2 assesses whether the findings transfer across model families. The evaluation covers both in-distribution (ID) and out-of-distribution (OOD) settings and examines whether compact adaptation provides consistent gains over strong pretrained checkpoints. Pretrained foundation models outperform the implemented seasonal naive and XGBoost baselines across the evaluated aggregation levels. On MOIRAI Large, rank-16 LoRA reduces normalized mean absolute error (NMAE) by 7.5% on average while updating 3.146 million parameters, about 1% of the model, and retains most of the improvement from full fine-tuning. Chronos-2 achieves lower pretrained NMAE than MOIRAI but benefits less from adaptation. OOD results show that adaptation benefits vary by target, with the clearest gains for the local-area series and several buildings. Overall, LoRA provides a favorable accuracy-efficiency trade-off when adaptation yields a meaningful gain over a strong pretrained checkpoint.