Review of Boiler Intelligence: From In-Furnace Sensing to Decision Optimization

Driven by global carbon neutrality targets, coal-fired power generation is undergoing substantial operational changes. Modern boilers must operate more flexibly while maintaining low emissions and reliable performance. This transition has increased the need for intelligent monitoring, operator-supervised optimization, and safety-constrained decision support. Using a reproducible search and screening procedure, this review examines the development of boiler intelligence across four interconnected technological stages. At the sensing layer, data-driven soft sensors support rapid prediction of flue gas emissions, while graph-structured spatiotemporal models are used to characterize flame and combustion states. At the modeling layer, physics-informed neural networks (PINNs) and proper orthogonal decomposition reduced-order models (POD-ROMs) are reviewed as routes for accelerating physical-field reconstruction. Surrogate models coupling computational fluid dynamics (CFD) with artificial intelligence (AI) provide another route to rapid prediction and can incorporate physical constraints. These fast field models can also serve as components of boiler digital twins for online assessment and operational guidance. They may also support early warning when abnormal conditions emerge. At the decision layer, reinforcement learning, model predictive control, and multi-objective optimization are reviewed for combustion and selective catalytic reduction (SCR) control. Because these applications are safety-critical, autonomous control must remain within actuator limits and established operating margins. Emission requirements and ammonia-slip constraints must also be satisfied. Safe deployment further requires fallback mechanisms, cybersecurity protection, and human supervision. Industrial application is still limited by data scarcity and lifecycle concept drift, while limited interpretability and simulator-to-real transfer create additional challenges. Edge latency and insufficient validation under abnormal conditions remain important barriers. Finally, industrial foundation models and large language models are discussed mainly as knowledge interfaces and operator-assistance tools rather than direct safety-critical controllers.