A Novel Explainable Impedance Identification Method Based on Deep Learning for the Vehicle-Grid System of High-Speed Railways

As many electric multiple units (EMUs) integrate into the traction network, low-frequency oscillation (LFO) issues may occur in railway systems. Impedance-based frequency-domain stability analysis is a common method used to analyze the stability of vehicle-grid systems. However, the vehicle-grid system is a complex single-phase system with numerous nonlinear components, making it difficult to establish accurate and online-analyzable impedance models. To address these challenges, this article proposes the residual feedforward neural network (ResFNN) suitable for EMU impedance identification. The ResFNN integrates residual connections into the deep feedforward neural network (FNN), which can improve the model prediction accuracy and get rid of the complex model derivation process. To explain the contribution of the input parameter to the model, the SHapley Additive exPlanation (SHAP) method is adopted in this article. Furthermore, according to the above contributions, this article proposes an unequal step size data collection method to optimize the operating point (OP) parameter step size and obtain a small but high-quality dataset. Then, combined with the vehicle-side and grid-side impedance models, this article can realize online vehicle-grid system stability analysis. Finally, actual case studies are conducted with multiple vehicles connected to the vehicle-grid system to validate the feasibility and accuracy of online stability analysis.