Dynamic combustion optimization of a pulverized coal boiler considering the wall temperature constraints: A deep reinforcement learning-based framework

Abstract

Coal-fired power generation boilers are susceptible to low combustion efficiency, pollutant exceedance, and over-temperature tube burst during deep peak shaving due to load fluctuation and mass flow rate variation, which seriously affects unit economics, environmental protection, and safety. In this paper, the issue of boiler combustion optimization under variable load conditions is investigated using a data-driven approach. (1) Based on the published literature, a multi-objective prediction model based on Channel Selection Convolutional Neural Network (CS-CNN) for boiler thermal efficiency, NOx emission and wall temperature was developed, taking into account the safety of the unit under fast variable loads (the range of wall temperature varies widely and is prone to over-temperature). (2) Conventional multi-objective optimization algorithms are time-consuming and hard to satisfy the real-time decision-making of combustion in the furnace. In this research, we developed a combustion decision-making agent based on the Twin Delayed Depth Deterministic Policy Gradient (TD3). By interacting with the predictive model and learning the combustion strategy, the agent has the ability to optimize the unfamiliar operating conditions of the boiler. Simulation experimental tests were carried out on real historical data from a 600 MW down-fired boiler and the results showed that thermal efficiency increased by 0.411 %, and NOx emissions decreased by 17.701 mg/m3, with the safety of the wall temperature maintained. A single decision by the TD3 is almost instantaneous, with an average time of 0.004 s. The decision-making efficiency was much better than the traditional search-based optimization algorithms.