A Causal EWT-LSTM Framework for Anomaly Detection and Localized Reconstruction of Indoor Temperature Time Series in District Heating Buildings

Indoor temperature time series in district-heating buildings are often contaminated by anomalies embedded in nonstationary, multiscale thermal dynamics. This study proposes a hybrid Empirical Wavelet Transform and Long Short-Term Memory (EWT-LSTM) framework for adaptive anomaly detection and localized reconstruction. Evaluated on 15 min interval data from 45 residential units over a 112-day heating season, the framework operates via a highest-frequency branch for anomaly detection and a full-modal branch for signal repair. Quantitative results show that the EWT Highest-Frequency LSTM (EWT(HF)-LSTM) achieved the best anomaly discrimination among decomposition variants with an average F1-score of 0.531. For anomaly repair, the full EWT-LSTM produced the highest fidelity with a localized Root Mean Square Error (RMSEa) of 0.818 °C. Furthermore, thermal comfort validation demonstrated that EWT-LSTM successfully prevented the severe comfort degradation of up to −82% in Exceeded Degree-Hours caused by unstable Empirical Mode Decomposition (EMD)-based reconstructions. These concrete results confirm that the proposed framework effectively provides clean, physically coherent temperature data for downstream district heating operations.