Deep Learning Framework for Estimating Energy Consumption by ALSTM Network: A Case Study of Hotel Buildings

With the rise of urbanization and growing demand for high-quality indoor environments, building energy consumption has increased significantly, particularly during seasonal extremes. Accurate forecasting of building energy use is essential for optimizing operational strategies and enabling demand-side flexibility. This study investigates the energy consumption of six hotel buildings in China using supply-side data from March and August, employing a novel hybrid framework that integrates Hilbert–Huang Transform (HHT), and an Attention-based Long Short-Term Memory (ALSTM) network. The proposed model effectively captures non-linear, non-stationary load dynamics by decomposing signals into intrinsic mode functions and applying attention mechanisms to emphasize critical temporal patterns such as morning ramp-ups and evening peaks. Empirical results demonstrate that the ALSTM achieves MAE = 91.21, RMSE = 114.52, CV-RMSE = 19.8%, and R² = 0.97, corresponding to a 64.6% reduction in RMSE compared to ANN and a 3% reduction compared to LSTM. These improvements confirm the superiority of the hybrid approach (ANN < LSTM < ALSTM) for handling complex seasonal and temporal variations in building energy demand. While current evaluation is limited to a single climatic zone, the findings highlight the potential of the proposed framework as a scalable, accurate, and interpretable tool for energy forecasting, with future work focusing on deployment across diverse building typologies and climates.