Voltage Gain Prediction Method for Buck Converters in Sneak Circuit based on Graph Neural Networks

Under the condition of high-frequency operation, the Buck converter is affected by the parasitic parameters of the MOSFETs and is prone to trigger the sneak circuit behavior, resulting in the variation of the working mode and the deviation of the voltage gain characteristics from the ideal situation. Traditional mathematical modeling methods are difficult to uniformly describe the nonlinear dynamic process dominated by parasitic parameters, and have limited adaptability to changes in working conditions, which restricts their application effect in engineering design. In this paper, a topological equivalent model of the Buck converter with parasitic parameters is established based on the Bond Graph method. The device characteristics and connection relationships are encoded through the graph structure form, and a voltage gain prediction model suitable for the regression training of graph neural networks (GNN) is constructed. The model structure adopts a combination of multilayer graph convolutional neural network (GCN) and fully connected regression network, combined with the multi-condition data generated by PSIM simulation, covering the DCM case and two types of sneak cases, and is trained for different parasitic parameter configurations. Experiments show that this model can achieve high-precision prediction under all working conditions, effectively reflect the gain characteristics of the converter caused by parasitic parameters, and is suitable for dynamic characteristic modeling and parameter optimization under highfrequency operating conditions.