Machine Learning and Time-Series Foundation Models for Day-Ahead Electricity Forecasting in a University Building
Day-ahead electricity forecasts can help building operators schedule heating, ventilation, and other services around anticipated demand. Choosing a forecasting model also requires balancing accuracy with its forecast-time inputs, local training, and computational requirements.
This study evaluates machine learning and time-series foundation models for day-ahead electricity demand forecasting in a university building, comparing their accuracy, computational efficiency, and practical applicability.
What happened
Researchers benchmarked statistical, machine learning, and pretrained time-series models (TimesFM, XGBoost, LSTM) for electricity forecasting, using chronological validation across 57 test days. They analyzed performance differences across workdays, weekends, and holidays, evaluating the impact of weather covariates and computational requirements.
Why it matters
Accurate day-ahead electricity forecasting enables optimized building energy management, reducing costs and improving resource allocation. The study provides practical insights for selecting models that balance accuracy, training requirements, and operational constraints.
- TimesFM with covariates provides the best overall forecasting accuracy
- XGBoost offers faster computation for local training scenarios
- Weather data improves holiday forecasts but may reduce accuracy for regular days
- Pretrained models eliminate the need for building-specific training
- LSTM models show diminishing returns compared to simpler approaches
The bigger picture
The findings establish pretrained TimesFM as a strong baseline for building energy forecasting without requiring domain-specific training, while XGBoost remains a viable alternative for applications prioritizing rapid local adaptation.
The study demonstrates that pretrained time-series models like TimesFM provide significant forecasting improvements over traditional methods, while XGBoost remains a competitive option for applications requiring faster local adaptation. The results highlight the trade-offs between model complexity, computational requirements, and forecast accuracy in energy management systems.
Technical details
- TimesFM 3.0 (pretrained time-series foundation model)
- XGBoost (gradient-boosted decision trees)
- LSTM (long short-term memory networks)
- Chronological validation for time-series data
- Weather covariates as external input features
- Mean absolute error (MAE) as primary evaluation metric
FAQ
What is the primary application of the models tested?
Day-ahead electricity demand forecasting for building energy management systems
How does TimesFM compare to XGBoost in terms of performance?
TimesFM achieves lower overall error rates but requires more computational resources, while XGBoost offers faster local training with comparable accuracy in some cases
What role do weather covariates play in the models?
Weather data improves holiday forecasts but slightly reduces accuracy for workdays and weekends, with no significant impact on overall mean absolute error
Why is LSTM less effective than other models?
LSTM requires longer training times without demonstrating superior performance compared to TimesFM or XGBoost
What are the key benefits of using pretrained models like TimesFM?
They eliminate the need for building-specific training while providing strong baseline accuracy for electricity forecasting tasks
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by arXiv cs.LG on Oct 12, 2026.
Comments
Sign in to join the discussion.