Empirical Assessment of Time-Series Foundation Models For Power System Forecasting Applications
TimesFM and Chronos Bolt go head-to-head on ERCOT solar, wind, and load data
A new empirical study from University of Colorado Boulder researchers Muhy Eddin Za'ter and Bri-Mathias Hodge provides the first comprehensive benchmark of time-series foundation models for power system forecasting. Using the high-resolution ARPA-E PERFORM dataset from the Electric Reliability Council of Texas (ERCOT) grid, the team evaluated 10 models—including Google's TimesFM, Amazon's Chronos Bolt, Salesforce's MoiraiL, MOMENT, Tiny Time Mixer, Temporal Fusion Transformer, PatchTST, TimeXer, LSTM, and CNN—across eight core capabilities. These include zero-shot performance, fine-tuning efficiency, multivariate input/output handling, horizon sensitivity, generalization to unseen sites, probabilistic forecasting, and context window effects.
The study offers clear guidance for grid operators: foundation models like TimesFM and Chronos Bolt shine in zero-shot and data-scarce scenarios, reducing the need for extensive historical data. However, simpler deep learning baselines like LSTM and CNN remain more practical for standard forecasting tasks with sufficient data. The findings are critical for renewable energy integration, as accurate solar and wind forecasting directly impacts grid reliability and cost efficiency. The paper is available on arXiv (2604.22077) and will be presented at the 2026 IEEE Power & Energy Society General Meeting.
- 10 models benchmarked including TimesFM, Chronos Bolt, MoiraiL, and MOMENT
- 8 capabilities assessed: zero-shot, fine-tuning, multivariate, horizon sensitivity, generalization, probabilistic forecasting, and context window effects
- Foundation models excel in data-scarce scenarios; LSTM/CNN remain competitive for standard tasks
Why It Matters
Grid operators get clear guidance on when to use AI foundation models vs. traditional methods for renewable forecasting.