Latent-Feature-Informed Neural ODE Modeling for Lightweight Stability Evaluation of Black-box Grid-Tied Inverters

TL;DR

This work tackles the challenge of stability assessment for black-box grid-tied inverters where internal models are unknown or proprietary. It introduces Latent-Feature-Informed Neural ODE (LFI-NODE), a continuous-time neural ODE that learns a single intrinsic vector field $f(x,u;\theta)$ while leveraging latent perturbation features through a Jacobian $J_{\mathrm{ref}}$ to regularize small-signal dynamics. The method demonstrates data efficiency, achieving accurate trajectory and eigenvalue predictions with as few as 48 short trajectories and outperforming discrete baselines by one to two orders of magnitude in key metrics. By unifying large- and small-signal modeling under a physics-informed learning framework, LFI-NODE offers a practical pathway for high-fidelity, lightweight stability assessment of GTIs and lays the groundwork for scaling to system-level stability with Neural DAE extensions.

Abstract

Stability evaluation of black-box grid-tied inverters is vital for grid reliability, yet identification techniques are both data-hungry and blocked by proprietary internals. {To solve this, this letter proposes a latent-feature-informed neural ordinary differential equation (LFI-NODE) modeling method that can achieve lightweight stability evaluation directly from trajectory data.} LFI-NODE parameterizes the entire system ODE with a single continuous-time neural network, allowing each new sample to refine a unified global model. It faithfully captures nonlinear large-signal dynamics to preserve uniform predictive accuracy as the inverter transitions between operating points. Meanwhile, latent perturbation features distilled from every trajectory steer the learning process and concurrently reveal the small-signal eigenstructure essential for rigorous stability analysis. Validated on a grid-forming inverter, {The LFI-NODE requires one to two orders of magnitude fewer training samples compared with traditional methods, collected from short time-domain trajectories instead of extensive frequency-domain measurements.} {Furthermore, the LFI-NODE requires only 48 short transients to achieve a trajectory prediction error at the hundredth level and an eigenvalue estimation error at the tenth level, outperforming benchmark methods by one to two orders of magnitude.} This makes LFI-NODE a practical and lightweight approach for achieving high-fidelity stability assessment of complex black-box power-electronic systems.

Figures (8)

• Figure 1: The comparison between the universal-intrinsic-modeling paradigm (red) and multi-local-modeling paradigm (blue).
• Figure 2: Schematic diagram of the neural network structure and learning process of the LFI-NODE approach.
• Figure 3: Hardware platform and evaluation process for the GFM inverter case
• Figure 4: HIL simulator results for verification scenario. (a). represents input condition 1. (b). represents input condition 2.
• Figure 5: Comparison of large-signal behavior and small-signal behavior of the proposed LFI-NODE and NODE, NARX methods under different external conditions. (a), (b) Results of the output voltage Vd. (c), (d) Results of the output current Iq. (e), (f) Eigenvalue results at the stabilization point. (g), (h) Enlarged plot of the eigenvalue near the zero point. The left column (a, c, e, g) represents input condition 1, and the right column (b, d, f, h) represents input condition 2.