Enhancing hierarchical learning of real-time optimization and model predictive control for operational performance

In process control, the integration of Real-Time Optimization (RTO) and Model Predictive Control (MPC) enables the system to achieve optimal control over both long-term and short-term horizons, thereby enhancing operational efficiency and economic performance. However, this integration still faces several challenges. In the two-layer structure, the upper layer RTO involves solving nonlinear programming problems with significant computational complexity, making it difficult to obtain feasible solutions in real-time within the limited optimization horizon. Simultaneously, the lower layer MPC must solve rolling optimization problems within a constrained time frame, placing higher demands on real-time performance. Additionally, uncertainties in the system affect both optimization and control performance. To address these issues, this paper proposes a noval hierarchical learning approach for RTO and MPC controller using reinforcement learning. This method learns the optimal strategies for RTO and MPC across different time scales, effectively mitigating the high computational costs associated with online computations. Through reward design and experience replay during the hierarchical learning process, efficient training of the upper and lower layer strategies is achieved. Offline training under various uncertainty scenarios, combined with online learning, effectively reduces performance degradation due to model uncertainties. The proposed approach demonstrates excellent performance in two representative chemical engineering case studies.