Evaluating Time-Series Foundation Models for Cooling Demand Forecasting with Little Data

Modern buildings increasingly integrate local energy generation, consumption, and storage, often involving multiple energy carriers such as electricity and thermal energy. This complexity creates opportunities for cost-efficient operation based on accurate forecasts of key energy parameters such as the cooling demand. However, generating such forecasts on a building-level can be challenging, as individual buildings likely exhibit highly specific consumption patterns and often offer only limited historical data. Time Series Foundation Models (TSFMs) offer a promising solution to this problem due to their ability to generalise across forecasting tasks and adapt to new domains with minimal data via fine-tuning. This study evaluates three state-of-the-art TSFMs (MOIRAI, MOIRAI-MoE and Chronos) in both zero-shot and fine-tuned settings. The models are tested on real-world energy consumption data from a split-type cooling unit used in a server room, spanning a period of less than four months. Outdoor air temperature is included as a covariate to assess its impact on prediction accuracy. Results are compared to two baseline models. Our findings show that Chronos, when incorporating outdoor air temperature, achieves a substantial improvement in forecasting accuracy for three-day ahead forecasts, reducing the Mean Absolute Percentage Error (MAPE) to as low as 2.57 %. In contrast, MOIRAI and MOIRAI-MoE show no significant benefit from the inclusion of temperature information. Overall, the study demonstrates that TSFMs represent a promising alternative for cooling demand forecasting in Section 5. Section 6 summarises the findings and outlines future research directions.