A novel twin time series network for building energy consumption predicting.

Energy consumption prediction in buildings is crucial for optimizing energy management. The latest research faces three critical challenges: (1) Insufficient temporal correlation extraction and prediction accuracy, hindering widespread adoption and application; (2) The positive impact of timestamp embedding in time series prediction under multi-mode decomposition; and (3) The issue of adaptive coupling with multi-source data. To overcome these issues, the study proposes Twin Time-Series Networks (T2SNET), which incorporates a time-embedding layer and a Temporal Convolutional Network (TCN) to extract patterns from Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), along with an adaptive fusion gate to combine energy consumption and meteorological data. The model was evaluated on datasets from university dormitories, office buildings, and school classrooms, showing significant improvements over the optimal baseline method. For instance, on the university classroom dataset, T2SNET reduced MAE by 4.56%, RMSE by 9.45%, and MAPE by 3.16% compared to the CEEMDAN-RF-LSTM model. These results highlight T2SNET's effectiveness in predicting building energy consumption, providing a robust solution for energy management systems. The proposed method, along with baseline model code and data, has been updated and is available at https://github.com/HaileiYuan/T2SNET-Pro.git.