Toward precision and intelligence: The convergence and evolution of artificial intelligence, data science, and control theory in building HVAC load forecasting

Accurate prediction of heating, ventilation, and air conditioning (HVAC) load demands has become a cornerstone technology for achieving building energy efficiency, enhancing power grid stability, and advancing carbon neutrality objectives in the built environment. This comprehensive review systematically examines the transformative evolution of HVAC load prediction methodologies over the past decade, with particular emphasis on the convergence of data science, artificial intelligence (AI) algorithms, and modern control theory. Through critical analysis of 96 peer-reviewed publications, this study traces the progression from traditional statistical approaches to sophisticated hybrid frameworks integrating advanced data preprocessing, machine learning algorithms, and intelligent control strategies. The review reveals significant accuracy improvements, with Mean Absolute Percentage Errors declining from over 15% in early traditional methods to approximately 2–5% under real-world conditions and below 1% in simulation-based environments, although cross-study comparisons require caution due to heterogeneous datasets and validation protocols. Key findings demonstrate that data preprocessing and feature engineering serve as critical enabling foundations whose quality often determines model performance potential, while the systematic integration of signal decomposition with hybrid deep learning architectures constitutes the primary driver of accuracy improvements. The integration of prediction capabilities with Model Predictive Control and reinforcement learning has demonstrated 10–30% energy savings with simultaneous comfort improvements. Persistent challenges including generalization limitations, uncertainty quantification deficits, and interpretability concerns are identified, alongside emerging opportunities in physics-informed AI, Large Language Model integration, and federated learning, all at early maturity stages for HVAC applications. This review uniquely contributes by: (I) establishing data preprocessing as a critical enabling foundation through systematic examination of signal decomposition and feature engineering; (II) tracing the complete algorithmic evolution from traditional methods through hybrid deep learning frameworks; (III) examining prediction-control integration with Model Predictive Control and deep reinforcement learning; and (IV) critically assessing emerging paradigms with explicit maturity evaluations. The future of building energy management lies in autonomous, adaptive systems combining physics-based understanding with AI adaptability, optimized control with uncertainty-aware decision-making, and individual building intelligence with community-scale coordination.