Accurate household electricity forecasting has become essential as power systems integrate distributed renewable generation and storage at scale. A new benchmarking study compared 30 state-of-the-art forecasting models to identify which approaches best predict residential load demand, rooftop solar output, and battery storage operation across different time horizons and geographic regions.
Researchers evaluated statistical methods, machine learning algorithms, deep neural networks, and emerging time series foundation models (TSFMs) using real-world datasets from Belgium, Germany, the Netherlands, and the United Kingdom. They applied standardized data preprocessing and consistent evaluation metrics to ensure fair comparison. Results showed LightGBM, a gradient-boosted machine learning model, consistently outperformed other trained approaches in accuracy and reliability across most forecasting tasks.
However, the study identified a notable advantage in newer transformer-based foundation models: zero-shot forecasting capability. Timer-XL and similar TSFMs demonstrated strong performance without task-specific retraining, potentially reducing deployment time and computational overhead for utilities facing new forecasting challenges. This capability becomes increasingly valuable as distributed energy resources proliferate and forecasting requirements expand.
The research highlights important trade-offs. Traditional statistical models remain interpretable and lightweight, suitable for resource-constrained environments. Machine learning methods balance accuracy and computational cost effectively. Deep learning approaches require substantial data but can capture complex temporal patterns. Foundation models offer flexibility but demand careful validation before deployment.
For grid operators and utilities, the findings suggest a portfolio approach: leveraging LightGBM for high-accuracy load and solar forecasting at 15-minute to hourly intervals, while exploring zero-shot foundation models for emerging use cases and rapid deployment scenarios. Performance varies notably across forecast horizons and regions, emphasizing the importance of local testing before model selection.
The publicly released source code and standardized benchmarking framework enable utilities to conduct local evaluations with their own data, accelerating adoption of optimal forecasting strategies tailored to specific operational needs.
