Volume 61

Integrating Deep Learning with Modelica for a COâ‚‚ Heat Pump System: A Hybrid Modeling Case Study in Oslo Ge Song, Qian Zhang, Natasa Nord

Abstract

Modelica is a powerful language for modeling thermal systems. However, the heat pump models available in open source Modelica libraries often lack sufficient accuracy for CO₂ heat pump systems. Unlike conventional heat pumps that use condensers, CO₂ heat pumps employ gas coolers, introducing significant thermodynamic differences not well captured in standard libraries. While commercial libraries offer more accurate CO₂ heat pump models, they are often costly, computationally intensive, and require substantial development effort—especially when the research focus is on the overall heating system rather than the detailed behavior of the heat pump itself. This study presents a deep learning approach to develop a CO₂ heat pump model based on real system data. The trained model was integrated into a Modelica-based thermal system simulation, enabling a hybrid modeling framework. A case study using a school building located in Oslo demonstrated the effectiveness of the approach. The hybrid deep learning-based heat pump model, when combined with the Modelica system model, improved overall simulation accuracy compared to a traditional Modelica-only model. Furthermore, the proposed framework is flexible and can be applied to other industrial thermal systems, offering a scalable solution for data-driven thermal system modeling.

Keywords Deep learning, Modelica, COâ‚‚ heat pump, thermal system, Hybrid modeling