An Adaptive Transfer-Learning-Assisted Modular Learning Approach for Control-Oriented Building Modelling

Machine learning (ML)-based Model Predictive Control (MPC) has shown promise for smart buildings, with online ML model adaptation often integrated to enhance long-term performance.However, abrupt building thermodynamic changes, such as opening/closing windows, can hardly been captured by the conventional online ML model adaptation.An Adaptive Transfer-Learning-Assisted Modular Learning (ATML) framework combined with dataset truncation is proposed by this paper to address the challenge.A simulation case study using a residential building was conducted for evaluation.Results show that, compared with conventional online ML model adaptation, ATML reduces prediction errors by 32.4% after 2 hours of the systemic change and by 25.4% after 22 hours, thereby demonstrating its effectiveness in coping with abrupt systemic changes.