Continual learning for building energy forecasting via hybrid elastic weight consolidation and synaptic intelligence

Accurate building energy forecasting is crucial for smart grid management and carbon reduction, but conventional static models suffer from performance degradation due to non-stationary environments and concept drift. To address this catastrophic forgetting challenge, this study proposes a dynamic energy forecasting framework based on a hybrid continual learning approach. The model combines an asymmetric encoder-decoder convolutional gated recurrent unit backbone with an attention mechanism to extract complex spatiotemporal features. Crucially, a dual-regularization mechanism that merges Elastic Weight Consolidation (EWC) and Synaptic Intelligence (SI) is introduced to balance network stability and plasticity without requiring storage of historical data. Empirical evaluations across four real-world public building datasets demonstrate that the framework consistently outperforms static baselines and vanilla regularization implementations. The proposed approach achieves high forecasting accuracy, with an average R² of 0.94 and a mean absolute percentage error of 5% under the optimal monthly update configuration, while also yielding a 78% reduction in maximum forecasting errors during severe environmental transitions. This research establishes a robust replay-free regularization methodology for continuous, adaptive energy management in long-term dynamic operations.