EWCM-MPC: An Adaptive Multimode Process Control Based on Continual Learning and Model Predictive Control

Due to changes in production conditions and strategies, industrial processes have various operation modes. However, the model based on a single operation mode cannot adapt to multimode process, leading to a degradation in the performance of model predictive control (MPC). Furthermore, most multimode process MPC require establishing a separate predictive model for each operation mode, causing the effectiveness of MPC to highly depend on the switching strategy. To address these issues, this paper proposes a continual learning-based MPC method for multimode process. First, an autoencoder and a predictor are investigated as the deep learning networks, and the continual learning strategy based on Elastic Weight Consolidation (EWC) is used to update model among various processes, thereby a single predictive model that can automatically represent multimode process is established. Then, the single predictive model is integrated with the MPC framework, and precise control of the multimode process is achieved by solving optimization problem. Finally, numerical simulation experiments are designed to demonstrate that the proposed method achieves minimal overshoot and high accuracy in controlling multimode process.