13 April 2025 in Arabian Journal for Science and Engineering
2 April 2026 in SSRN Electronic Journal
Accurate forecasting of building energy consumption is critical for smart energy management and achieving carbon neutrality targets. However, conventional static models have limited adaptability to dynamic environmental shifts, such as meteorological fluctuations, equipment degradation, and evolving occupancy occupancy patterns, leading to long-term performance degradation. To address this, this study proposes a dynamic energy forecasting model based on a hybrid continual learning framework. By integrating Elastic Weight Consolidation (EWC) and Synaptic Intelligence (SI) regularization terms, the model preserves historical knowledge while adapting to new data distributions without storing raw data, effectively mitigating catastrophic forgetting. The model adopts an encoder-decoder architecture with Convolutional Gated Recurrent Units (dubbed EDCG) as its backbone, where ConvGRUs extract complex spatiotemporal features. Meanwhile, the decoder combines a Gated Recurrent Unit (GRU) with an attention mechanism to capture critical temporal dependencies and the nonlinear weather-energy coupling relationship.Furthermore, Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is employed to preprocess energy sequences, enhancing robustness against data volatility. Experimental results on datasets from real-world public buildings demonstrate that the proposed method achieves an R² of 0.97 and a MAPE of 3.3%. The framework significantly outperforms traditional EWC, Accumulative Retraining (AR), and static models in accuracy, training efficiency, and stability, validating its utility in non-stationary environments.