Event-Structured Physics-Informed Neural Networks for Differentiable Critical Clearing Boundaries

What the math gives to ML

The paper's transferable contribution is not the power-system model itself, but the event-aligned treatment of a piecewise dynamical system. Instead of learning one trajectory across discontinuous changes in the governing equations, it uses a separate phase representation and passes the terminal state of one phase exactly into the next, avoiding interface-penalty errors. This structure can be transplanted to neural ODEs, state-space models, and physics-informed surrogates for systems with known mode switches, interventions, contacts, or solver events. The most practical first test is to replace a single continuous-time MLP surrogate with chained phase modules and compare stability, accuracy near events, and training sensitivity to interface loss weights.