A Noise Immune Physics Informed Data-Driven Digital Twin for Inverter Component Degradation Estimation

Performance deviation of inverters in grid-connected distributed energy resource systems, due to component degradation from environmental stressors, is a major reliability concern. This work presents a data-driven digital twin (DT) of an H-bridge inverter with output filtering, developed using physics-informed machine learning (PIML). A ‘switching’ digital model is created using experimental data collected from the inverter prototype and optimized via particle swarm optimization. Later, continuous degradation is introduced to filter parameters, including parasitic resistances and the aging/reliability factor in filter components, generating a comprehensive dataset for training supervised ML models, for fast, accurate degradation estimation. To account for real-world conditions, measurement noise—such as ripple and ground loop effects—is introduced for capacitor parasitic resistance. In this scenario, the proposed PIML-DT approach, enhanced with a modified loss function and physical constraints, accurately estimates capacitor equivalent series resistance despite partial and noisy measurement data. It achieves an R 2 of 0.6918 and RMSE of 0.0053, demonstrating strong resilience to noise and effective tracking of inverter degradation. This PIML-DT approach supports reliable, data-driven condition monitoring and long-term health assessment of power electronic converters.