Efficient Control Parameters Optimization for Grid-Connected Converter via Data-Driven Surrogate Model

Conventional methods for optimizing power converter control parameters primarily rely on extensive simulations and exhibit limited adaptability across diverse operational scenarios. To overcome these challenges, this study introduces a novel approach with deep neural network (DNN)-based surrogate model for efficient parameter optimization. The proposed method is implemented in a classical grid-connected converter. During the preprocessing stage, a high-accuracy classifier with 97% validation accuracy is incorporated to reduce the scale of the dataset. The model is trained using 16,000 validated simulation cases from a real-time digital simulator (RTDS), covering eight representative scenarios. The DNN structure is carefully designed to balance computational efficiency with generalization capability. The trained surrogate model enables performance evaluation in under one millisecond. Particle swarm optimization (PSO) is then used to identify optimal control parameters in less than 10 seconds while considering multiple scenarios simultaneously. Simulation results verify the accuracy of the surrogate model in performance prediction. The results indicate that multi-scenario optimization meets all constraints, while single-scenario optimization does not. Additionally, hardware-in-the-loop experimental tests confirm the effectiveness of the selected control parameters.