Study Of Parallel Computing System For Model Predictive Control

Model predictive control (MPC) is one of the advanced control algorithms. It utilizes an explicit process model and measurable outputs to predict the future response of a plant and minimize the difference between the plant's response and the reference response. In this thesis, a lot of research has been done on the key technologies of MPC systems based on the project "MPC system for fuel cell" which is collaborated with the Global Research Center (Shanghai) of General Electric. Firstly, the MPC design methods are investigated. Based on linear matrix inequality theory, a low conservative state feedback robust MPC design method is proposed. The dynamic output feedback robust MPC design method is also developed via the combination of generalized eigenvalue optimization and instrumental matrix transformation. In order to improve the computational performance of the real-time MPC systems, an FPGA based double-precision floating-point parallel matrix multiplier is proposed. Since the performance of a single FPGA cannot meet the system requirements of the real-time MPC systems under most circumstance, a master-slave distributed multi-FPGA parallel computing system and the corresponding algorithm which is suitable for MPC systems are proposed.In chapter 1, a comprehensive description of the significance of the research work is given and the current research status of MPC, FPGA and parallel computing are surveyed. The research orientation and architecture of this thesis are also described.In chapter 2, the improved state feedback MPC design method with low conservation is proposed. The dynamic output feedback robust MPC design method is developed based on generalized eigenvalue optimization and instrumental matrix transformation. The robust MPC for a class of uncertain systems with saturation actuator is also investigated.In chapter 3, the computation of MPC is analyzed and the computation load is decided. Based on the analysis results, an FPGA based double-precision floating-point matrix multiplier is investigated and proposed. This multiplier integrates P~2 processing elements (PEs) in a single FPGA and matrix multiplication is implemented through the parallel computing of the PE array. A data pre-processing module is configured before the PE array which can avoid the computation of 0 element blocks of sparse matrices, and thus the performance of sparse matrix multiplication is improved. The relationship between performance of the multiplier and the number of PEs, frequency and bandwidth is also discussed.In chapter 4, a master-slave distributed multi-FPGA matrix multiplication parallel computing system suitable for MPC is investigated and proposed. The processing element of the system ultilizes the FPGA based matrix multiplier. RapidIO or Ethernet is adopted as the system interconnection network with star-connected topology. The system has certain advantages such as low communication overhead and low requirement of hardware resources. Based on the characteristic of MPC computation, a suitable parallel algorithm is proposed.In chapter 5, the design schemes proposed in chapter 3 and chapter 4 are implemented and the prototype system is designed. Based on the experimental results of the system, a parallel computing system for proton exchange membrane fuel cell MPC is designed.The last chapter concludes the main research work of this thesis. The prospect of the future research is also presented.