LLM-Enhanced Physics-Informed Dynamic Adaptive Multimodal Forecasting Framework for NOx Prediction in CFB Boilers

Abstract

Circulating fluidized bed (CFB) coal-fired boilers play a key role in global energy production, but their nitrogen oxides (NOx) emissions have brought significant environmental challenges. Accurate prediction of NOx content is very important for effective emission control. The existing methods generally have defects such as insufficient prediction accuracy and unable to adapt to changes of boiler operating conditions. This paper proposes a Large Language Model (LLM)-enhanced physics-informed dynamic adaptive multimodal forecasting framework for NOx prediction in CFB boilers. The framework is designed as a task-specific industrial prediction model, where LLM component serves as a semantic encoder for unstructured data and structured process variables are encoded by temporal networks. A dynamic context window is introduced to adjust the input sequence length according to operating stability. Temporal and semantic attention mechanisms are further combined to fuse sensor-based dynamic features with log-based semantic features. Soft physics-guided regularization terms associated with thermal NOx and fuel NOx formation are introduced to improve physical consistency. Online adaptation and uncertainty quantification are further incorporated to enhance robustness and provide confidence information. Experiments on real industrial data collected from a 300 MW CFB boiler unit show that the proposed framework achieves lower prediction errors and more stable performance than the compared baseline models under different operating conditions. Note to Practitioners—This article is motivated by the need for reliable NOx emission prediction in CFB boilers under practical operating conditions. The proposed framework is intended to provide an online decision-support tool for boiler operators and emission control engineers. It combines structured process variables with unstructured operation records, uses a dynamic context window to adapt to changing operating stability, and introduces soft physics-guided regularization to reduce physically unreasonable predictions. The uncertainty estimation module further provides confidence information, which can be useful when the predicted NOx concentration approaches an emission limit. The method can be implemented using routinely collected DCS data and operation-log records. In deployment, reliable data preprocessing and periodic updating are important for stable performance. The framework is not designed to replace existing combustion control or denitration control systems, but to provide predictive information that can support earlier operational adjustment and more reliable emission monitoring in CFB boiler operation.