Model Predictive Control Of Flying Capacitor Multilevel Converters

In recent years,industries such as rail transportation,data centers,industrial manufacturing,and power systems have posed higher demands on power conversion technology.Developing new power electronic converters with high power density,efficiency,high voltage levels,and medium to high power capacity is of great significance.However,the physical characteristics of semiconductor devices constrain the operating performance of power converters.To improve the operating voltage and power level of power converters,the academic community has proposed multi-level converters.Among these,the flying capacitor multi-level converter(FCMC)has attracted more and more attention due to its high power density and strong expandability.However,the application of FCMC converters in high-power situations is still constrained by two factors.Firstly,FCMC has many internal state variables,making it complex to control,especially in situations with higher levels where appropriate control methods are lacking to maintain stable FCMC capacitor voltages.Secondly,as the number of levels in FCMC increases,more voltage sampling circuits are inevitably introduced,leading to an increase in system volume and cost.This paper takes the demand for high-performance control strategies for FCMC as the starting point and combines FCMC topology with model predictive control(MPC)to design predictive models and optimization functions based on the characteristics of the converter.This can significantly optimize its steady-state performance,simplify control complexity while leveraging the good dynamic performance of predictive control.Additionally,this paper proposes a voltage estimation method applicable to FCMC,which can accurately and rapidly obtain the voltage values on each capacitor without using sensor circuits.The research focus and major innovative work of this paper are reflected in the following aspects:(1)In response to the challenge of the exponential increase in MPC computation complexity with the increase in the number of levels,a new concept of neural network-model predictive control(ANN-MPC)is proposed for the first time.This method uses neural networks to replace the control process of FCS-MPC,and the heavy computational burden in traditional FCS-MPC is shifted from online iteration to the offline training process of ANN,effectively reducing the control process’ s computational resource usage by the ANN-MPC controller.Comparative experimental results in a flying capacitor five-level inverter show that ANN-MPC can reduce FPGA resource utilization by more than 50% while providing the same control performance as FCS-MPC.(2)In response to the problem of variable switching frequency and high ripple in circuits frequently encountered in FCS-MPC,a novel fixed-frequency modulation model predictive control applicable to multi-level structures is proposed.The vector analysis-based modulation model predictive control proposed in this paper retains FCS-MPC’s outstanding transient performance while ensuring the steady-state performance of the converter is consistent with the carrier phase-shift strategy.In other words,it unifies the transient performance of traditional predictive control and the steady-state performance under PWM modulation.Moreover,it no longer depends on the selection of cost functions and weight coefficients.(3)In response to the common problem of model predictive control(MPC)sensitivity to system parameter mismatch,this paper proposes an adaptive model predictive control based on neural networks(ANN-AMPC).This method uses a neural network model to learn the FCMC operating data under different system parameters.To achieve the ability to track changes in system parameters,the control signal of the previous working cycle is also added to the input vector of the ANN.After training,the ANN-AMPC can adaptively track changes in system parameters and achieve optimal control performance without requiring an additional system parameter identification section.Moreover,since the convergence of ANN under different model parameters is completed offline,ANN-AMPC can provide better dynamic performance.(4)In response to the unfavorable factors of exponential growth in FCS-MPC with the increase of FCMC levels,variable output waveform frequency,and large output ripple,a new CCS-MPC control strategy is proposed-Distributed Model Predictive Control(DMPC).DMPC is based on the Nash equilibrium strategy and has a distributed structure,and the optimal duty cycle is obtained through chain iteration.The most significant advantage of DMPC over FCS-MPC is that it can achieve fixed-frequency control.With the increase of FCMC levels,only a few sub-processor modules need to be added,without the need to redesign the entire converter,ensuring the linearization of the growth rate of computational burden and facilitating fast and fixed-frequency control of high-level FCMC.With its advantages in control performance and algorithm implementation cost,DMPC provides powerful technical support for the promotion of FCMC to higher levels and power levels.(5)The flying-capacitor multi-level topology requires additional sampling circuits to adjust the flying-capacitor voltage,which will lead to an increase in system cost and volume.To address this issue,this chapter studied the fast capacitor voltage observation technology of FCMC topology and proposed two capacitor voltage estimation methods,namely SGDE based on backpropagation algorithm and ANNE based on ANN.The SGDE strategy does not require offline data and can online approximate the capacitor voltage value through real-time sampled system states.It has simple computation and strong scalability.This paper verified the effectiveness of SGDE through simulation and experiments in five-level and seven-level FCMCs.The simulation and experimental results show that the operating frequency of SGDE can reach 100 k Hz,and the estimation accuracy is higher than 99%.The ANNE strategy first collects operational data of the multi-level converter,and then offline trains a neural network to estimate the flying-capacitor voltage value online.This section verified the effectiveness of ANNE through simulation and experiments in a five-level FCMC.The simulation and experimental results show that the operating frequency of ANNE can reach 100 k Hz,and the estimation accuracy is higher than 99%.Moreover,because the convergence process of the neural network has been completed offline,ANNE has strong disturbance resistance in online estimation process.