Research on Dynamic Prediction Model for Building Energy Consumption Based on EWC-SI

Building energy consumption accounts for a significant portion of national energy use, and accurate prediction is essential to meet “dual carbon” goals. Traditional static models are computationally expensive and lack real-time adaptability, while dynamic models often struggle with balancing efficiency and accuracy. This paper proposes a continual learning model based on Elastic Weight Consolidation (EWC) and Synaptic Intelligence (SI)—termed EWC-SI—that leverages dual regularization to assess parameter importance, adapt to dynamic changes, and avoid catastrophic forgetting. Experiments on public building data demonstrate that the EWC-SI framework, underpinned by an encoder-decoder architecture, reduces mean absolute error (MAE) by 67% and shortens training time by 40% compared to LSTM and traditional static methods.