Symbolic regression for data-efficient cooling load forecasting in real-world HVAC systems

Accurate short-term cooling load forecasting is essential for efficient HVAC operation in buildings, particularly in mission-critical facilities such as hospitals. However, many modern forecasting tools depend on large volumes of historical data and complex black-box algorithms, making them difficult to deploy in real systems and unsuitable for cold-start environments where only limited recent data are available. This study investigates symbolic regression (SR) as a transparent and lightweight alternative for short-term cooling load prediction under data-scarce conditions. Using real operational data from a hospital cooling plant in Vancouver, models were trained on window lengths ranging from 3 to 16 days and evaluated across five in- domain and out-of-distribution periods spanning summer and shoulder seasons. A reliability-driven selection strategy was introduced to identify robust symbolic expressions across multiple stochastic runs, favoring stable performance and physical plausibility over peak accuracy. Results show that symbolic regression can extract consistent, interpretable relationships primarily driven by outdoor temperature and time-of-day effects, even with as little as three to five days of training data. The selected SR models achieved strong forecasting accuracy and generalization, outperforming tuned machine-learning baselines trained on the same limited data in this case study. SR-5 achieved 17.65% MAPE compared to 22.14% for XGBoost-Tuned (20% reduction) on in- domain validation, with up to 67% reduction under out-of-distribution conditions. Diebold-Mariano tests confirmed that these improvements were statistically significant. Beyond forecasting accuracy, SR models demonstrated exceptional deployability: the best equation required only a handful of floating-point operations, negligible memory, and delivered near-instant predictions suitable for microcontroller-level hardware. These findings show that symbolic regression offers a practical and transparent pathway for rapid deployment of building energy forecasting, supporting edge-based control, explainability, and fast commissioning. Future work will extend evaluation to year-round operation, multiple building types, and embedded field testing.