PAFE-TQNet: a period-aware temporal query network with temporal difference consistency loss for building electricity load forecasting

Accurate building electricity load forecasting is essential for energy-efficient building operation, demand-side management, and intelligent energy scheduling. Although existing time series forecasting models can capture temporal dependencies from historical observations, they often rely on backbone networks to implicitly learn periodic patterns and pay limited attention to daily load variations, hour-profile deviations, and the trend consistency of predicted curves. To address these issues, this paper proposes a Period-Aware Feature-Enhanced Temporal Query Network with Temporal Difference Consistency Loss, named PAFE-TQNet with TDCL. Based on the original TQNet framework, the proposed method introduces a period-aware feature enhancement module, PAFE-v3, to explicitly enrich the input representation using daily difference features and hour-profile residual features. Specifically, the daily difference branch describes the load variation between the current time step and the same hour of the previous day, while the hour-profile residual branch captures deviations from the typical intra-day load profile. In addition, a temporal difference consistency loss is designed to constrain the difference sequence between adjacent predicted values and the corresponding ground truth, enabling the model to improve both point-wise accuracy and trend consistency. Experiments are conducted on the BDG2 building electricity load dataset under multiple prediction horizons. The results show that the proposed method reduces the average MSE from 0.523170 to 0.520756 and the average MAE from 0.429420 to 0.426246 compared with the original TQNet baseline. These results demonstrate that the proposed method can provide a stable improvement for short- and medium-term building electricity load forecasting and has practical value for intelligent building energy management.