Powergrid Simulation with Graph Neural Networks
Electric power transmission networks operate under extremely high currents and voltages, making them susceptible to overload and stability problems. To ensure secure and reliable operation, system operators perform computationally intensive real-time simulations that evaluate risk and inform mitigation strategies. Conventional approaches typically employ nonlinear, non-convex solvers (such as Newton–Raphson methods or interior-point algorithms (IPOPT)) whose runtimes impose practical limits on scenario exploration. For example, a single-timestamp powerflow simulation of the French grid requires approximately 100 ms. In day ahead security assessments, forecasting over a 24-hour horizon refreshed every 15 minutes involves simulating up to 9000 contingencies (90% of the 10,000 possible node failures) within each quarter-hour window. As energy transition uncertainties drive finer time discretizations, potentially to 5-minute intervals, the same 5-minute interval permits only <3000 contingency evaluations, leaving the majority of critical failure modes unexamined. Consequently, the execution time of the solver emerges as a bottleneck that constrains both the breadth and depth of the risk analysis. With power grid topologies and operational regimes becoming increasingly complex, there is an urgent need for faster but equally reliable simulation techniques capable of supporting comprehensive contingency screening in near real time.