Abstract
Real-time dynamic simulation requires that each simulated time interval complete within the equivalent wall-clock CPU time, a strict budget that large power system models, with hundreds of thousands of state variables, cannot meet. Existing model reduction methods either sacrifice internal network detail, are valid only near a fixed operating point, or cannot adapt to changing contingencies online. This paper proposes an Adaptive Model Selection (AMS) framework that automatically reduces a full-detail dynamic model to a smaller, contingency-specific model by identifying which substations require full fidelity and simplifying the rest. Given the pre-disturbance operating state and event location, a pair of Graph Attention Network (GAT)-based ordinal predictors classify each component’s expected dynamic activity level; substations predicted to be inactive are simplified from node-breaker to bus-breaker representation. Applied to a realistic French transmission network model (over 6,000 buses), AMS reduces continuous and discrete model variables by approximately 60%, achieves up to 2.6x simulation speedup, and maintains trajectory errors below the numerical solver tolerance, demonstrating a practical route toward real-time operation of large-scale dynamic simulation.
Stefanos Eleftheriadis
PhD Candidate @ CUT
Mohammad Hashemnezhad
PhD Candidate @ CUT
Savvas Panagi
PhD Candidate @ CUT
