🤖 AI Summary
Short-term household electricity load forecasting demands practical zero-shot generalization without task-specific training. Method: This study presents the first systematic evaluation of time-series foundation models—Chronos, TimesFM, and LagLlama—in this setting, benchmarking their zero-shot performance against a from-scratch trained Transformer under identical experimental conditions and evaluating rigorously using multi-scale error metrics (MAE/MSE). Contribution/Results: All foundation models achieve performance on par with task-specific models, with TimesFM attaining the best zero-shot accuracy under large input windows—reducing MAE by 12.3% relative to the strongest customized model. Critically, none require gradient updates or domain-specific fine-tuning; all operate effectively with only minimal historical data. The results demonstrate that time-series foundation models are both effective and deployment-efficient for fine-grained, low-resource load forecasting scenarios.
📝 Abstract
Accurate household electricity short-term load forecasting (STLF) is key to future and sustainable energy systems. While various studies have analyzed statistical, machine learning, or deep learning approaches for household electricity STLF, recently proposed time series foundation models such as Chronos, TimesFM, or LagLlama have not yet been considered for household electricity STLF. These models are trained on a vast amount of time series data and are able to forecast time series without explicit task-specific training (zero-shot learning). In this study, we benchmark the forecasting capabilities of time series foundation models compared to Trained-from-Scratch (TFS) Transformer-based approaches. Our results suggest that foundation models perform comparably to TFS Transformer models, while the TimesFM foundation model outperforms all TFS models when the input size increases. At the same time, they require less effort, as they need no domain-specific training and only limited contextual data for inference.