Compositional Power System Dynamic Model Using Causality-Aware Physics-Informed DeepONet

Component surrogates that are accurate under given terminal voltage signals fail after being applied to a new network. We propose a one-shot, full-horizon, MILP-composition approach to overcome this gap by combining physics-informed deep operator networks with causality for fourth-order synchronous generators. The operator is trained on a 39-bus network with different line open topologies and various branch parameters, then evaluated on held-out networks. For the generalization capability test in 5400 held-out samples, with the true voltage trajectory supplied, $1.68{\times}10^{-2}$ p.u. (real) and $7.36{\times}10^{-3}$ p.u. (imaginary) of MAE for current are achieved. The voltage trajectories are unknown unless a time-domain simulation (TDS) is conducted, so an MILP is formulated to couple voltage-current through AC network equations and operator's input and output. With the true voltage trajectory not supplied, $2.04{\times}10^{-2}$ p.u. (real) and $7.86{\times}10^{-3}$ p.u. (imaginary) of MAE for current are obtained.