A novel data-driven approach for coal-fired boiler under deep peak shaving to predict and optimize NOx emission and heat exchange performance.

Coal-fired boilers considering deep peak shaving operate under wide and variable load requirements with complex and varying working conditions, making the prediction and optimization of combustion target variables challenging. This study proposed a data-driven approach for the prediction model building, tuning and combustion optimization. In this approach, a neural network for prediction based on Gaussian kernel and long short-term memory (LSTM), which captured and fused the working condition and temporal information, was first proposed. To address incomplete historical data coverage of working conditions, a fine-tuning framework with the idea of transfer and few-shot tuning was then introduced. Finally, an improved fitness function considering both optimization objectives and combustion adjustment limitations was designed to mitigate the risks associated with large-scale adjustments of operable variables. By conducting experiments on a 1000 MW boiler, targeting NOx concentration and the outlet temperature of the air pre-heater, the proposed method demonstrated high prediction accuracy with R 2 values of 0.979 and 0.989, and significantly reduced mean squared error (MSE) by 98.05% and 97.37% on new conditions compared to untuned models. It achieved the reductions in NOx and temperature of 24.70 mg m − 3 and 2.013 ∘ C , and decreased the average changes of operable variables by 9.66% and 11.24% compared to using only rigid constraints. [Display omitted] • Proposed a Gaussian kernel and LSTM structure to capture and fuse information. • Introduced the idea of transfer and few-shot tuning for combustion prediction models. • Validated the fine-tuning framework for online combustion model tuning. • Provided soft constraints on variable adjustments using an improved fitness function. [ABSTRACT FROM AUTHOR]