Sequential Learning-Based Physics-Informed Neural Network Modeling for Parameter Estimation on a Power Converter
This paper tackles offline parameter identification for a non-ideal DC-DC buck converter using Physics-Informed Neural Networks (PINNs). The core idea is a staged training schedule that progressively unfreezes physical parameters—inductance, capacitance, and parasitic resistances—to improve convergence and interpretability. The work matters because accurate component-level estimates enable condition monitoring and predictive maintenance without intrusive frequency-response tests.
The paper presents a well-motivated but incremental application of sequential PINN training to power converter parameter estimation. While the staged strategy yields sensible convergence trajectories and moderate accuracy for $L$, $C$, and $r_L$, claims of outperforming a simultaneous baseline rest almost entirely on the abstract without supporting figures or tables in the main text. The manuscript also contains internal contradictions: Table III-A lists the release order as L, then C, then $r_L$, then $r_C$, whereas the Conclusion re-labels the five-stage schedule as L, then $r_L$, then C, then $r_C$. These inconsistencies weaken confidence in the empirical protocol.
The staged training heuristic is intuitively justified by sensitivity analysis and produces clean, monotonic parameter trajectories. The authors explicitly model parasitic resistances in the governing ODEs, capturing damping and efficiency effects absent from idealized textbook models. The $8^4$ factorial sweep and a fixed random seed add methodological structure. As noted, the Stage 1 data-only surrogate integrates forward with no substantial drift, establishing a solid baseline before physics penalties are introduced.
Three issues stand out. First, the paper promises an order-of-magnitude improvement over simultaneous training in the abstract, yet no simultaneous baseline results appear in the results section, making the claim unverified. Second, the proposed unfreezing order is inconsistent: Table III-A explicitly releases L first, then C, then $r_L$, then $r_C$, whereas the Conclusion asserts the five-stage schedule followed L, then $r_L$, then C, then $r_C$, calling that sequence the natural hierarchy of parameter influence. Third, the robustness sweep is ambiguously described: it is unclear whether the 4,096 factorial cases involve re-estimation or forward evaluation of a fixed post-training model, and the claim that progressively determined parameters are generalizable to unseen operating points is unsupported if the network was not retrained per case.
The internal evidence supports moderate claims about staged convergence but falters on comparative assertions. The RMSE heatmaps and trajectory overlays in Fig. 7 and Fig. 8 show predictable behavior—inductance dominates voltage error, while leakage resistance leaves a negligible signature—but the manuscript never quantifies the simultaneous-training baseline it invokes. Related work is cited adequately, yet the paper does not position its accuracy against comparable PINN or Kalman-filter studies from the literature with matched experimental conditions.
Reproducibility is hampered by the absence of published code, data, or solver scripts. Although Table III lists hyperparameters, network depth and width are never stated, and the ODE solver used to generate synthetic ground truth is unspecified. Gradient clipping, learning-rate staging, and z-score scaling are documented, but without the architecture or layer counts, independent reproduction would require substantial guesswork.
A staged physics-informed neural network (PINN) is proposed for identifying the inductance, capacitance and parasitic resistances of a DC buck converter from time-domain data. Gradients are released in the order of L to C to rL to rC, a sequence derived from sensitivity analysis. The five-stage schedule recovers L and C to within 5% and the series ESR to 10%, while a three-stage variant attains the same accuracy 11% faster. An 84 factorial sweep confirms that the trained PINN remains accurate for simultaneous 8% perturbations of all four components, achieving a waveform RMSE of 2.6 x 10-2 over 5 ms-an order-of-magnitude better than a simultaneous-training baseline.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.