Incremental online learning with model augmentation for AI-enabled optimization of building cooling systems

With the increasing adoption of AI-enabled optimization in building cooling systems, online learning has been widely explored to improve model adaptability under component performance degradation. However, its practical deployment in real-time control remains challenging due to limited online data, measurement quality, and the risk of catastrophic forgetting. This study proposes an incremental online learning strategy based on a residual-enhanced surrogate model to enable robust and stable model adaptation. The strategy combines a pre-trained hybrid surrogate model with a residual learning branch, where only the residual parameters are updated during operation, thereby preserving previously learned knowledge while allowing gradual adaptation to evolving conditions. A conservative learning-rate Adam optimizer is adopted to ensure stable parameter updates. Validated through hardware-in-the-loop testing on a physical smart control station, the proposed strategy restores model accuracy under component performance degradation, with R2 improving from approximately 0.4 to above 0.9. Representative error metrics also decrease markedly, with the relative error reduced from 6.77% to 0.95%. This improvement enhances optimization performance, increasing the energy saving rate from 1.12% to 5.77%, while also providing improved robustness under measurement noise.