Predictive Scheduling of Wet Flue Gas Desulfurization System based on Reinforcement Learning

Abstract With the development of renewable energy, loads of thermal power units fluctuate, resulting in the trade-off between the frequent switching of auxiliary equipment and the economic-emission benefits in the wet flue gas desulfurization (WFGD) system. In this paper, the predictive scheduling problem is formalized, considering the power consumption, emission punishment and the switching frequency of slurry circulation pumps with finite prediction sequence of load and sulfur in coal. Model-free off-policy reinforcement learning (RL) is applied to solve the unclear and drifting system dynamic. Considering a real system, the framework setting and an emulator is introduced. Compared with traditional scheduling policies and the case without prediction, the proposed framework shows obvious advantages in terms of comprehensive performance and approximates the theoretical optimal solution at the steady-state. Moreover, the policy keeps the performance by adapting to the drifting without manual intervention, which demonstrates a broad application prospect in similar scenarios.