Multi-scale interpretable temporal prediction network for building energy consumption forecasting

Accurate forecasting of building energy consumption is crucial for optimizing energy management, reducing operational costs, and achieving carbon neutrality goals. This study proposes a multi-scale interpretable temporal prediction network model (ITSFN), which enhances prediction accuracy and reliability through the collaborative optimization of long short-term temporal (LSTM) networks and Kolmogorov-Arnold networks (KAN). The model integrates temporal-environmental feature decoupling with a dynamic attention mechanism, explicitly decomposing time-series data into seasonal, trend, and residual components to construct a structured feature space. It employs a parallel architecture of gated recurrent units (GRU) and multi-head attention to model multi-scale features. Tested on an energy consumption dataset from a university building in a hot-summer/cold-winter region, ITSFN outperforms traditional models: it reduces the root mean square error (RMSE) of total energy consumption prediction by 13.9% compared to LSTM and decreases the RMSE of sub-item energy consumption prediction by 31.1% compared to Transformer. Additionally, ITSFN enhances the noise suppression coefficient to 0.89 through feature decoupling, achieves a local attention angle of 0.92 in mutation regions, and reduces over-smoothing by 29.6% compared to traditional methods. By quantifying feature contributions, the model reveals the evolutionary patterns of component weights, further validating its effectiveness and practical applicability.