• An adaptive modular hybrid modeling approach for a building-integrated GSHP-DH system. • A physics-guided, event-triggered mechanism enables real-time adaptation to changes. • An Accuracy Recovery Gradient (ARG) metric quantifies the model’s adaptability rate. • White-box and black-box models were compared across modules for accuracy and runtime. • The physics-guided all-black-box configuration delivered the best overall performance. Accurate and efficient building energy models with high adaptability are essential for online energy management of multi-energy heating systems in complex buildings. Although various modeling approaches were utilized to balance accuracy and computational efficiency, they remain limited in flexibility, interpretability, and adaptability to changing control conditions. This study develops an adaptive modular hybrid modeling approach for a cold-climate, multi-purpose building supplied by a hybrid ground-source heat pump (GSHP) and district heating (DH) system. Within the proposed framework, white-box and black-box models were developed for each component using measured data and tested in all-white-box, all-black-box, and hybrid configurations to evaluate trade-offs between accuracy and computational time. A physics-guided, event-triggered adaptation mechanism was introduced and validated with measurements from real-world scenarios representing varying operating conditions to ensure real-time adaptability. Adaptability was quantified using the novel accuracy recovery gradient factor (ARG), which demonstrated fast and effective performance recovery after disruptions. The all-black-box model, guided by the building white-box model for synchronization, achieved the best overall performance. It estimated hourly electrical consumption with a CVRMSE of 7.5% while running over 1000 times faster than the hybrid configuration. It could also automatically detect deviations, trigger adaptation, and regain acceptable accuracy in 28 and 170 minutes under minor and major changes in operating conditions, respectively. The results demonstrated the potential of the proposed modular hybrid model, supported by an event-triggered adaptation mechanism, for real-time energy management in complex multi-energy buildings.
Amini et al. (Sun,) studied this question.