A dynamic adaptive multi-source aggregated transfer learning framework for individual thermal comfort prediction

Abstract

Occupants’ thermal comfort is critical for building operation and energy management. However, thermal comfort is influenced by multiple factors and changes dynamically, making accurate prediction challenging. As a result, existing methods face key challenges, including limited accuracy, inefficient generalization, and lack of dynamic adaptability. To address them, this study develops a dynamic adaptive multi-source information aggregated transfer learning framework. Specifically, a multi-source information aggregated MLPMixer model (TCMLPMixer) is developed to improve prediction accuracy under the same input modalities and a comparable model size by modeling feature interactions among diverse thermal comfort factors. In addition, a similarity-aware efficient transfer learning (TL) strategy is introduced to improve transfer efficiency under different data-availability conditions. Furthermore, a dynamic adaptive online learning procedure is developed to enable non-intrusive pseudo-label acquisition, and to support adaptive fine-tuning with different TL strategies according to data availability, thereby facilitating dynamically adaptive thermal comfort prediction. An orthogonal experimental design is employed to construct a multi-source individual thermal comfort dataset. Experimental results show that our framework consistently outperforms baselines. Specifically, under the same input modalities and a comparable model size, the pre-trained TCMLPMixer outperforms the DNN model by 7.40 percentage points, while under limited-data conditions in the target domain, it outperforms the DNN-based transfer model by 4.82 percentage points in accuracy. Moreover, it exhibits strong dynamic adaptability through progressive online updates, with target-domain accuracy increasing from 79.63% to 92.57% during long-term deployment. These results demonstrate that our framework can achieve improved prediction accuracy, enhanced generalization, and dynamic adaptability, demonstrating its potential applicability to occupant-centered intelligent building control.