Aiming at the characteristics of multivariable coupling, pronounced nonlinearity, time-varying behavior, and time delay effects in clean coal ash content during dense medium cyclone separation, which make it difficult for traditional control methods to achieve high-precision and stable regulation, this paper proposes an intelligent control method based on the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm. First, a data-driven environment model based on Long Short-Term Memory (LSTM) is constructed using historical operational data from a coal preparation plant to enable offline training of the reinforcement learning policy. Second, the state space, action space, and multi-objective reward function are designed for the ash content control task. On this basis, the standard TD3 algorithm is improved by introducing a hierarchical experience replay mechanism to enhance the utilization efficiency of critical samples, and a gated feature attention enhancement network to strengthen state representation under complex operating conditions. Experimental results demonstrate that the proposed method achieves the best overall performance among the compared approaches, with a mean absolute error (MAE) of 0.1190 and a root mean square error (RMSE) of 0.1938. The compliance rates within the target intervals of ±0.2 and ±0.5 reach 83.29% and 97.14%, respectively. Compared with Model Predictive Control (MPC), the proposed method improves the compliance rate under strict constraints by approximately 9.58 percentage points, indicating superior fine control capability. In addition, the proposed method outperforms the benchmarks in terms of error distribution, fluctuation suppression, and steady-state maintenance. These results verify that the improved TD3 method can effectively enhance the accuracy and stability of clean coal ash content control, providing a feasible solution for intelligent optimization control of quality indicators in complex industrial processes.
Li et al. (Mon,) studied this question.