Research On Parameter Identification Of Power Converter Based On Neural Networks

Power converters usually undertake core power control and conversion work in the system,and have been widely used in aerospace,defense military,new energy generation,and household appliances and other fields.Since the power converter is in the high-frequency and high-voltage operating state for a long time,each device in the circuit will be degraded to different degrees.The parameter identification technology can monitor the health condition of the device,and can detect early faults and perform maintenance in time.Therefore,research on the parameter identification technology of power converter can provide technical support for prognosis and health management of power converter and improve system reliability.This thesis mainly studies the parameter identification technology of power converter based on neural network.The research content contains following aspects:(1)The failure mechanism of key components in the power converters is analyzed and their fault characteristic parameters are determined,and the non-ideal equivalent circuit model of key components and Buck converter are established in this thesis.And the monitoring signals with low test cost are selected by sensitivity analysis and correlation analysis,which reduces the redundancy of test signals and lays a foundation for power converter parameter identification.(2)The parameter identification method of power converter based on convolutional neural network is studied.The training mechanism of auxiliary regression generative adversarial nets is improved.And the flow of parameter identification method based on two algorithms is studied.(3)The experimental research on parameter identification of Buck converter based on neural network is carried out.Aiming at the problem that the signal characteristics are difficult to extract,the voltage time domain signals are used as the convolutional neural network input for automatic feature extraction.And the parameter identification model based on the impored wasserstein generative adversarial mechanism is established to identify the characteristic parameter value of the key components in the main circuit.The simulation experiment analyzes the difference of the identification performance between the proposed method and the commonly used neural network algorithm model under different conditions,and validates the effectiveness of the proposed method in the Buck converter experimental test platform.And the relative error of the identification of each device parameter is basically maintained below 5%.