Physics-informed machine learning for power system dynamics: A framework incorporating trustworthiness

In this paper, we introduce a framework for trustworthy physics-informed machine learning (PIML) applied to power system dynamics, emphasizing transparency, interpretability, and explainability of Neural Network (NN) models for safety-critical applications. With the massive deployment of converter-interfaced resources and the growing uncertainty in both generation and demand, there is an urgent need for computationally efficient models that can reliably approximate complex dynamic behaviors. By embedding physical laws within machine learning models, physics-informed NNs (PINNs) offer the potential for faster, more accurate dynamic simulations. We present model reduction and validation methods, introducing novel correctness verification techniques and an interpretability-driven architecture, Kolmogorov–Arnold Networks (KANs), to enhance trust in NN approximations. Further, we put forward PINNSim, an integrated simulation tool leveraging PINNs for large time-step dynamic simulations, demonstrating its performance on reduced-order synchronous machine (SM) models. This work contributes to the design of transparent, interpretable, and rigorously validated PIML frameworks to accelerate the deployment of reliable AI tools in power systems.