AI-driven modeling and short-term forecasting of thermal energy transfer in district heating systems

Abstract

Accurate monitoring of thermal energy transfer (ΔE) is essential for efficient district heating operation, particularly when direct flow-based measurements are unavailable or incomplete. This study evaluates artificial intelligence (AI)-based soft sensing and one-hour-ahead forecasting of ΔE using temperature-only input features from five heating seasons of supervisory control and data acquisition data. The analysis separates full-dataset modeling from active-heating (ON-only) operation and distinguishes instantaneous estimation from short-term forecasting to isolate zero-inflation effects and assess true dynamic predictability. For instantaneous estimation, tree-based ensemble models achieved high accuracy under ON-only conditions, with Extreme gradient boosting (XGBoost) and Random Forest obtaining R2 values of 0.909 and 0.894, respectively. In contrast, one-hour-ahead forecasting was more challenging, especially under ON-only operation, where XGBoost and Random Forest achieved R2 values of 0.567 and 0.543, respectively. These results demonstrate a clear distinction between estimating current thermal behavior and predicting its short-term evolution. The findings show that temperature-based AI soft sensors can reliably estimate real-time thermal energy transfer, whereas short-horizon forecasting remains constrained by dynamic variability, unobserved operational factors, and active-load fluctuations.