Machine Learning-Based Forward Design Approach for Grid-Connected Inverters With Enhanced Stability and Dynamic Performance

Grid-connected inverter (GCI) is extensively utilized in renewable energy power systems. However, these systems are prone to cascaded instability when connected to the power grid, even if both the inverter and the grid are stable themselves. To achieve an integrated design that considers cascaded stability and dynamic response, this article proposes a forward design method for GCI based on machine learning, aiming to design the control parameters including phase-locked loop and current loop controllers. First, based on the theoretical model of few samples, the impedance feasible region is rapidly trained by improved K-nearest neighbor and random forest (RF) algorithms, as well as the preliminary feasible samples are obtained. Second, the neighborhood of preliminary feasible samples is searched based on the trained impedance feasible region, thus further obtaining the parameter feasible region. Compared with the theoretical model-driven method, the proposed method achieves rapid design by few samples, which has low complexity and fast computation speed. Finally, the experimental verifications are carried out on a 5 kW inverter prototype, validating the effectiveness of the proposed design method.