A hybrid deep learning approach for building energy prediction with STL decomposition

Accurate building energy consumption prediction is crucial for optimizing energy efficiency and operational flexibility in modern buildings. However, existing prediction methods struggle to handle the highly complex, nonlinear, and multifactor-influenced nature of building energy data, often failing to fully exploit the intrinsic value and underlying characteristics of varying data components. To address these limitations, this study aims to develop a novel hybrid deep-learning framework that enhances multi-step prediction accuracy by decoupling and independently modeling the underlying patterns of energy time-series data. Methodologically, the proposed approach integrates STL with a divide-and-conquer predictive strategy. First, the original energy consumption series is decomposed into three interpretable components: trend, seasonality, and residual. Next, the random forest algorithm is applied to identify the most relevant input variables from historical energy, calendar, and meteorological data for each specific component. Subsequently, tailored deep-learning models are assigned based on component characteristics: LSTM networks model the long-term trend, the Transformer architecture captures the periodic seasonal patterns, and a Res-LSTM tackles the highly stochastic residual component. The proposed framework is validated using real-world data from the public BDG2 dataset. Experimental results demonstrate that the STL-based hybrid model significantly outperforms single prediction models (including BPNN, SVR, RNN, LSTM, transformer, and Res-LSTM). Specifically, compared to the best-performing single model, the proposed method yields improvements of 20.1%, 12.7%, and 23.9% in mean absolute error (MAE), root mean square error (RMSE), and coefficient of variation of RMSE (CV-RMSE), respectively. With achieved final values of 0.0397 kW for RMSE and 0.94 for the coefficient of determination (R2), this study concludes that the proposed hybrid methodology effectively mitigates noise and non-stationarity in energy data. It provides a highly accurate, robust, and interpretable tool to support intelligent building energy management and system scheduling.