Data-driven thermal models for smart energy management in heating system

Forecasting the thermal behavior of flexible heating assets is essential for real-time energy management in heating systems. While thermal dynamics are complex, the optimization algorithms within energy management systems (EMS) require low-order models with minimal computational load to make rapid, real-time decisions. To bridge this gap, this study develops and validates low-order, data-driven models for a heat pump, an electric boiler, and the indoor temperature of an office building. These models are designed for integration into the Digital Twin EMS of an industrial site.

The electric boiler is characterized by its efficiency frequency distribution, which is centered at 85%, allowing it to be represented by a constant value. For the heat pump, a third-order polynomial captures how the coefficient of performance (COP) depends on the outdoor temperature, with a mean absolute error (MAE) of COP = 0.41. Indoor temperature dynamics are described with a discretized first-order model whose constant parameters are identified via four-minute nighttime regression; daytime disturbances are estimated either from training data or a three-day rolling profile.

The developed indoor-temperature dynamic model predicts office temperatures with an overall MAE below a 0.3°C threshold, which humans cannot perceive. The resulting low-order models are suitable for integration into model-predictive algorithms that schedule the operation of the heating system.