A Novel Learning-Based MPC Method via Basic-Residual Cooperative Model

摘要

This study proposes a novel model predictive control (MPC) method based on the basic-residual cooperative model. Compared to existing learning-based MPC methods that rely on a single network model as prediction models for either static feature capture or dynamic adaptation, which often result in insufficient adaptability or compromised computational efficiency, the proposed method integrates a dual-network architecture: a Long Short-Term Memory (LSTM) network to capture static system features, and a self-attention feed-forward neural network to adapt to dynamic aspects. The convergence and stability of the resulting control system are proven through theoretical analysis. The effectiveness of proposed method is validated through numerical simulations and experiments. Experimental results show that the proposed MPC method can reduce the prediction model’s root mean square error by about 70% compared to classical static model-based MPC and cuts computational time by about 30% compared to classical dynamic model-based MPC. The proposed method significantly enhances the model adaptability and computational efficiency of nonlinear dynamic systems, such as autonomous vehicles and robots.