Improving extrapolation of data-driven thermal response prediction for active phase change systems using temperature excavation

Accurate prediction of the dynamic thermal response of active phase change systems (APCS) is essential for control-oriented building energy applications, yet data-driven models often show poor extrapolation capability under sparse-data training conditions. To address this issue, this study develops an APCS-oriented temperature excavation (TE) framework that expands sparse temperature data along thermally relevant trajectories to improve extrapolation performance. A scaled APCS experimental platform was used to collect minute-level temperature data, and nine machine learning models were evaluated using sparse-data training, operating-condition extrapolation, and unseen control-strategy scenarios. The results show that TE consistently improves extrapolation performance, with particularly strong gains for shallow models. For example, TE-ELM reduced RMSE from 0.812 to 0.354, corresponding to a 56.4% reduction, while TE-SVM substantially improved robustness under unseen cooling-source start–stop transients. Additional analyses indicate that TE enhances thermally relevant state coverage while maintaining high consistency with the original data. These findings indicate that TE can provide a more reliable predictive basis for future control-oriented APCS applications, although its actual energy-saving benefits still require verification through closed-loop control experiments.