Pretrained TimesFM 3.0 cuts day-ahead building electricity error to 19.97 kWh, 43% below weekly seasonal-naive
Synopsis
On 57 day-ahead test days for a university teaching building, pretrained TimesFM 3.0 with covariates reached 19.97 kWh hourly mean absolute error, about 43% below the weekly seasonal-naive baseline, without training building-specific weights.
Interpretation
A pretrained time-series foundation model can deliver the lowest overall day-ahead electricity error without any building-specific weight training. In the earlier comparison for the same building, locally fitted Prophet and XGBoost were the strongest, and the recurrent network did not give the best reported accuracy. Eleven configurations were compared on the same 57 test days and 1,368 hourly targets; TimesFM 3.0 with covariates had the lowest mean absolute error at 19.97 kWh and the lowest root mean squared error at 30.26 kWh.
The model treats past electricity consumption as context and extrapolates the next 24 hours with its pretrained weights held fixed, rather than updating them for this building. Locally fitted routes retrain at every forecast origin, with median fitting times of 0.84 s for XGBoost and 99.05 s for the LSTM. The test covers 34 workdays and 23 non-workdays, including 12 weekend days and 11 holidays; TimesFM 3.0 with covariates reduced holiday root mean squared error from 53.74 to 39.47 kWh.
The gain from covariates concentrates on closures and holidays, while routine teaching-day error rises slightly. Target-only TimesFM 3.0 had a workday mean absolute error of 16.65 kWh, lower than the 18.85 kWh of the covariate version. Replacing KNMI observations with ERA5 reanalysis changed mean absolute error by less than 0.001 kWh, and replacing them with historical IFS increased it by about 0.17 kWh.
Perspective
The findings apply to day-ahead hourly electricity forecasting for a modern, well-insulated teaching building at the Delft campus of The Hague University of Applied Sciences, over a 57-day test from 2 February to 30 March 2024 containing 34 workdays and 23 non-workdays. A researcher or engineer working on building operations or energy forecasting can use it as a starting point for this building: begin with a target-only pretrained forecast, compare it with a locally fitted model, and add covariates where holiday accuracy justifies the extra input preparation.
The paired-day intervals include zero, so the 57 days do not establish a clear daily mean-absolute-error advantage for TimesFM over XGBoost. The extended historical evaluation uses shorter early contexts and settings selected from later validation data, making it exploratory; summer is covered only once. The historical IFS records lack per-value issue times, so the forecast sequence an operator could have obtained cannot be reconstructed. The evaluated TimesFM 3.0 weights are released for non-commercial, non-production research use.