Accelerating Model Predictive Control with Neural Network Optimizer

Model predictive control (MPC) is an advanced control method widely used in automotive systems due to its ability to effectively handle state and control constraints of multivariable systems. The highly nonlinear characteristics of vehicle dynamics, however, often result in long solving time. Applying neural networks (NNs) to the fast solving of MPC problems essentially involves solving ordinary differential equations (ODEs), which is precisely what NNs excel at. A neural network optimizer (NN optimizer), capable of harnessing both data and physics, therefore, is proposed to accelerate MPC solving. The NN optimizer can handle constraints and prediction information, and apply a set of parameters to map different states and disturbances to control inputs. During implementation, it involves a small amount of matrix multiplication and addition, leading to a very short solving time and guaranteed solving accuracy. To verify the effectiveness of the proposed fast solving method, it was compared to Acados, an internationally leading nonlinear MPC solver in terms of solving speed, in the context of predictive cruise control (PCC) problems. The results indicate that the NN optimizer has achieved an improvement in solving speed of over 70% compared to Acados, with no sacrifice in control performance.