Day-ahead electricity forecasts enable building operators to schedule heating, ventilation and other services ahead of expected demand. Selecting a forecasting model requires balancing accuracy, required forecast-time inputs, local training effort and computational cost. This study benchmarks statistical methods, machine‑learning models, a recurrent neural network (LSTM) and the pretrained TimesFM 3.0 model on a university building. Chronological validation is performed on the same 57‑day test set, and configurations are compared across workdays, weekends and holidays, as well as local fitting and inference time. Weather covariates are taken from retrospective observations.
TimesFM 3.0 with covariates achieves the lowest overall error, reducing mean absolute error (MAE) by about $43\%$ relative to a weekly seasonal‑naive baseline. XGBoost remains competitive, but daily pairwise comparisons do not show a clear MAE advantage for TimesFM. TimesFM also performs strongly when using only past electricity consumption. Adding covariates improves holiday forecasts but slightly degrades workday and weekend performance, leaving overall MAE almost unchanged. TimesFM requires no building‑specific weight training, whereas XGBoost offers rapid local fitting and inference. The evaluated LSTM takes longer to train without outperforming either approach.
An exploratory historical analysis reveals pronounced seasonal variation, and a separate weather‑source comparison yields only minor changes in the chosen TimesFM configuration’s accuracy. These findings support pretrained TimesFM as a strong initial forecasting baseline for this building, with XGBoost as a competitive locally‑fitted alternative. Covariates are particularly useful when improved holiday forecasts justify the extra input preparation.
Review: TimesFM’s zero‑training advantage and robust performance make it the go‑to choice for quick deployment, while XGBoost’s flexibility suits scenarios that benefit from local tuning. Leveraging weather covariates can boost reliability on specific holidays.