The air-conditioning loads of industrial buildings exhibit considerable nonlinearity, nonstationarity, and multi-scale variability due to the interaction between production processes and external weather conditions. This study proposes a three-stage hybrid forecasting framework that integrates Variational Mode Decomposition and Sample Entropy (VMD–SE) with a transformer–Bidirectional Long Short-Term Memory (BiLSTM) network architecture, optimized using Quantum Particle Swarm Optimization (QPSO). The VMD–SE module reconstructs the load signal by entropy-based component screening, significantly reducing high-frequency noise while maintaining essential dynamic features. The dual-path transformer–BiLSTM network concurrently captures long-term global dependencies and short-term local changes, facilitating an extensive temporal representation. By automatically adjusting its core parameters, QPSO achieves more efficient convergence, increased stability, and enhanced global search. Two real-world datasets obtained from industrial facilities in Hubei, China, reflecting distinct seasons and load levels, were utilized for validation using ablation and comparative experiments. The proposed model outperformed benchmark models such as SVR, GRU, and CNN–informer, achieving a coefficient of determination (R2) of 0.981 23 and a root mean square error of 0.181 32. The findings indicate that the proposed model is effective and versatile across various situations, offering both theoretical understanding and practical guidance for intelligent HVAC scheduling, dynamic demand response, and low-carbon energy management in industrial facilities. Our results show that the model is robust and flexible, providing a practical and theoretically grounded approach for intelligent HVAC scheduling, demand response, and industrial carbon reduction.
Wei et al. (Sun,) studied this question.