Adaptive transfer learning for household return water temperature prediction based on domain discrepancy metric

Abstract

Individual household temperature control in district heating systems is crucial for improving energy efficiency and comfort. However, the limited availability of indoor temperature monitoring in Chinese residential buildings constrains the implementation of individualized household control. To address this issue, this study proposes a household return water temperature prediction model based on transfer learning for indoor temperature regulation. By classifying households into groups based on thermal load characteristics, a base model is first trained on households with available indoor temperature data (source domain) within each group, and then transferred via transfer learning to predict for households without indoor temperature data (target domain) in the same group. The base model for return water temperature prediction can achieve an MAE of 0.28–0.66 °C and a MAPE below 2.1 %. In the domain adaptation framework, the ratio of heat consumption (QK) and the difference in heat consumption (ΔQ) between the source and target domains are incorporated as domain discrepancy metrics to enhance the transfer model's robustness. Three households with distinct distribution characteristics are selected as case studies. The proposed model yields an average MAE of 0.47 °C and an average MAPE of 1.44 %. Compared to the traditional station-level and building-level uniform return water temperature control methods for households, the proposed model reduces the relative error by 5.7 % and 9.13 %, respectively, effectively improving the accuracy of individualized control.