Optimizing Zinc Electrowinning Processes With Current Switching Via Deep Deterministic Policy Gradient Learning

Publication date: Available online 13 November 2019Source: NeurocomputingAuthor(s): Xiongtao Shi, Yonggang Li, Bei Sun, Honglei Xu, Chunhua Yang, Hongqiu ZhuAbstractThis paper proposes a model-free Deep Deterministic Policy Gradient (DDPG) learning controller for zinc electrowinning processes (ZEP) to save energy consumption during the current switching periods. To overcome the problems such as inaccurate modeling and various time delays, the proposed DDPG controller utilizes various control periods and parameters for different working conditions. Strategies such as action boundary setting, reward function definition, state normalization are applied to ensure its learning performance. Simulations and experiments show that the DDPG learning controller can significantly decrease energy consumption during the ZEP current switching periods. The optimal control policy will be learnt for different working conditions with only one group hyperparameters. Furthermore, the smoother control actions of the DDPG controller will improve the stability and reduce more energy consumption by comparing with traditional proportional-integral (PI) controller, model predictive control (MPC) and artificial experiences. The artificial intelligence-based optimal control framework brings both energy saving and intelligence to zinc manufacturing plants.
Source: Neurocomputing - Category: Neuroscience Source Type: research