The Impact of Covariates on Zero-Shot Building Energy Forecasting Using Chronos-2 Foundation Model

Foundation models for time series forecasting have recently been applied to energy prediction tasks, where they can produce accurate forecasts without task-specific training. This study investigates the impact of two types of covariates, meteorological variables and calendar-based day type indicators, on the forecasting performance of Chronos-2, a state-of-the-art foundation model, in building energy consumption prediction. Using real-world monitoring data from two non-residential buildings at the ENEA Research Centre in Portici, Italy, we systematically evaluate seven configurations combining past and future covariates across multiple observation window lengths (7–28 days). Future meteorological covariates are derived from historical weather forecasts rather than observed weather data, ensuring that the evaluation reflects realistic operational forecasting conditions. The results show that incorporating day type indicators as both past and future covariates consistently delivers the highest forecasting accuracy, reducing CV-RMSE from 14.58% for the covariate-free baseline to 10.41% with a 28-day observation window. A day-stratified analysis further reveals that these improvements are concentrated on regime transition days, for which recent load history alone provides limited information about the operating conditions of the day being forecast. By contrast, meteorological variables, whether obtained from weather forecasts or historical observations, yield only marginal performance gains, suggesting that calendar-driven operational schedules are the primary determinants of energy demand in the buildings considered. These findings provide practical guidance for deploying foundation models in real-world energy building management systems and show that covariate selection is a key determinant of forecasting performance.