Adaptive Model Predictive Control Using Artificial Neural Network for Power Converters Under Parameter Uncertainties

There has been an increasing interest in using model predictive control (MPC) for power electronic applications. However, the demand of accurate system model and the common issue of inevitable system parameter uncertainties hinders the practical adoption of this highly promising control technique. In this paper, a new adaptive MPC using an artificial neural network (termed ANN-AMPC) is proposed to overcome these barriers. A power converter with a supervisor MPC controller is first designed and operated using matched model parameters. An ANN is then built and trained offline using the operating data of the supervisor MPC controller. In addition to the input components of the supervisor MPC, its previous output control signals are also added to the input vector of ANN. Next, an actual FPGA-based adaptive MPC controller is designed using the trained ANN to control the actual operation of power converter in real time. The trained ANN-AMPC can adaptively track the variation of system parameters and achieve the optimal control performance without extra identification part of system parameters. The basic concept, ANN structure, off-line training method, and online operation of ANN-MPC are described in detail. The dynamic response of the ANN-AMPC and conventional MPC under step changes in load parameters are analyzed and compared. The ANN-AMPC concept is validated by both simulation and experimental results on a kW-class 3-phase voltage source inverter (VSI) as the case study. It is demonstrated that the FPGA-based ANN-AMPC controller offers better dynamic performance in the presence of uncertain parameters while utilizes reduced FPGA resource requirement compared with the observer based MPC controller.