Event-Based Data-Driven Adaptive Model Predictive Control for Nonlinear Dynamic Processes
Jian Sun, Xi Meng, Junfei Qiao
Abstract
Data-driven model predictive control (MPC) has been regarded as an attractive control technology for nonlinear dynamic processes. However, as a model-based approach, data-driven MPC usually suffers from an inaccurate model, process uncertainty, and calculation burden. To provide an efficient and accurate controller for nonlinear dynamic processes, an event-based data-driven adaptive MPC (EDAMPC) scheme is proposed. First, a prediction model employing a self-organizing long short-term memory (SOLSTM) neural network is developed to obtain a compact model structure and improve the generalization ability. The structure of the SOLSTM neural network is constructed dynamically by integrating the activity and significance of the neurons. Second, an error-triggered online learning mechanism is developed to update model parameters adaptively based on prediction errors. The nonlinear process dynamics can be captured in the presence of process uncertainty. Third, an event-based control strategy is introduced to reduce communication and computation resources. The convergence of the SOLSTM neural network and the nominal stability of the EDAMPC scheme is analyzed. Finally, the effectiveness and superiority of the EDAMPC scheme are demonstrated via a numerical case and an industrial application.