Research On The Model Predictive Control System Optimization And FPGA Implementation Of The High Frequency DC-DC Converter

Increasing switching frequency is an important technical path to reduce the volume of DCDC converters and increase power density.However,high switching frequency poses new challenges to the designs of the converter and control system.In terms of converter design,the main challenge lies in the heat dissipation caused by the increased switching loss.With the application of the soft-switching technology and wide-bandgap(WBG)devices such as silicon carbide(Si C)and gallium nitride(Ga N),the switching loss has been significantly reduced.As a result,the switching frequency of DC-DC converter has achieved as high as 100 MHz.In terms of control system design,the main challenge is that the complicated calculation of high performance control is difficult to be completed within such a short switching cycle.At present,analog control is the main control method for the high-frequency DC-DC converters,but the loop stability margin and response speed of the analog control are determined by hardware,and cannot be flexibly adjusted during operation.Therefore,when the operating conditions change significantly,it is necessary to make a compromise between dynamic performance and loop stability margin,and it is difficult to implement advanced control algorithms.For this reason,digital control that has the advantages of flexible programming,high reliability,simple debugging and maintenance,and good scalability,has great application potential in high switching frequency control.Limited by the constraint of high switching frequency,the current digital control algorithms of high-frequency DC-DC converters mainly are hysteresis control and proportional integral derivative(PID)control,which have limited control performance.Model Predictive Control(MPC),as a dynamic programming based on model information,can make full use of model information to flexibly design multiple objective functions and constraints,and optimize the control strategy according to the predicted process behavior in real-time rolling.As a result,MPC can take into account the overshoot,dynamic adjustment time and other performance indicators,and is the preferred solution to improve control performance.However,the online calculation of MPC is complicated and difficult to implement at high switching frequencies.Although some scholars have further proposed explicit MPC(EMPC)in which the control law can be solved offline,EMPC is only suitable for linear time-invariant(LTI)control plant.When applied to the nonlinear and time-variant DC-DC converters,a large number of EMPC laws need to be stored and a time-consuming online search process is required.Even with the high computing speed field programmable gate array(FPGA),the switching frequency still does not exceed 400 k Hz.Moreover,since the EMPC laws are derived from the converter model offline,its control performance depends on the accuracy of model parameters,so further adaptive design needs to be added to ensure the control performance,making it more difficult to be implemented at high switching frequencies.In order to realize EMPC at higher switching frequencies and make it adapt to the variations of circuit operating point and parameter,this thesis carries out research on the EMPC system optimization and its FPGA implementation.The FPGA implementation is considered from the design stage of the EMPC system to achieve comprehensive consideration of multiple factors such as control performance,computing burden and storage burden.The main research contents of this thesis are as follows:(1)Design and optimization of the state feedback linearization based EMPC system:EMPC can be solved offline,so the complicated real-time optimization problem is simplified to the storage and search problem of the control laws,which reduces the real-time calculation burden.However,since EMPC is only suitable for LTI control plant,when applied to nonlinear and time-variant converter,repeated designs need to be done under different operating points,which leads to a large number of control laws,and greatly increases the storage burden and search time.To this end,this thesis utilizes state feedback linearization technology to construct the equivalent control object of "new state variables containing operating point information +LTI state equation",and transfers the effects of operating point variation to the new state variables,so only a set of EMPC laws are needed.Moreover,a sliding mode observer is constructed to ensure the accuracy of the new state variables,thereby realizing the adaptation to the operating point variation.This EMPC system not only retains the high dynamic performance of EMPC and grants EMPC the ability to adapt operating point variation,but also shortens the total execution time,and is more suitable for high-frequency control;(2)Design and optimization of the neural network(NN)based EMPC system: the serial calculation of new state variables,control law search and inverse calculation of control parameter in(1)restricts the reduction of the online calculation and storage burden.For this reason,this thesis proposes the idea of using a NN to fit a large number of EMPC laws,a single hidden layer,multi-node “wide NN” is selected to ensure the parallelism.A dimension expansion method is proposed to avoid the confusion of the fitting relationship under different operation points,in this way,the operating point information is transferred to the input variables of the NN,and the ability to adapt operating point variation is naturally achieved.The trained NN only needs dozens of parameters to equivalently express a large number of EMPC laws,which greatly reduces the burden of real-time storage and search.In addition,the parallel structure of the “wide NN” is highly matched with the parallel computing capability of the FPGA,which is more conducive to high-frequency control;(3)Adaptive algorithm of the NN based EMPC system: in(2),the offline EMPC laws fitted by the NN is solved offline.When the actual circuit parameters and model parameters are mismatched,the control performance will inevitably be reduced.In this regard,this thesis further proposes a real-time training method of the NN controller to adapt the parameter variation.Different from the two traditional training methods that use the instantaneous error or the total error of the dynamic process for optimization,this thesis uses the error of a period of the dynamic process to train the NN,and proposes a segmented objective function that can effectively describe the dynamic performance.A training law to minimize the objective function(i.e,dynamic performance improvement)is given,which can be executed in parallel with NN control calculation,which avoids extra execution time while achieving dynamic performance improvement.Extending the NN controller from “offline fitting” to “online optimization” to automatically adapt to the circuit parameter variation;(4)FPGA implementation of the proposed two EMPC systems: to efficiently implement the proposed two EMPC systems in FPGA,this thesis analyzes the typical links such as “loop traversal judgment”,“NN forward calculation”,“NN parameter training” and“serial-parallel execution” in the control systems,and formulates the corresponding FPGA implementation schemes based on the principle of “space for time”.This thesis also presents the FPGA implementation schemes of the above-mentioned two EMPC systems by taking the classic buck converter as the control plant.A series of experimental verifications at up to 3MHz switching frequency are performed to demonstrate the feasibility,and a comprehensive evaluation is also given to show the comprehensive performance.Based on the above research,this thesis has formed a set of EMPC-based control system analysis,design and FPGA implementation methods,which not only retains the advantages of EMPC,but also has the adaptiveness to circuit operating point and parameter variations,thereby meeting the high control performance requirement of the high-frequency DC-DC converters.Experimental results fully verify the effectiveness of the proposed EMPC systems and their FPGA implementation schemes.As a result,this research can provide design basis and reference for the high-performance control of the high-frequency DC-DC converters.