Operational risk quantification of power grids using graph neural network surrogates of the DC optimal power flow

This work develops surrogate models of a decision-making algorithm (DC OPF) to expedite Monte Carlo (MC) sampling-based grid risk quantification. Sampling-based risk quantification allows explicit computation of the risk associated with a given probabilistic forecast of power demand and supply. However, it requires solving a large number of optimization (DC OPF) problems within a short time, which is computationally demanding. The computational burden is alleviated by developing graph neural network (GNN) surrogates, because GNNs are especially suitable to model graph-structured data. In contrast to previous works that developed GNN surrogates to predict bus-level (generator dispatch) decisions or line flow, we develop models to directly predict zonal/system level quantities needed for grid risk assessment. That is, in addition to generator dispatch and line flow, we develop GNN models that directly predict zonal or system level reserve shortage and load shedding. The benefits of these GNN surrogates are demonstrated using four synthetic grids (Case118, Case300, Case1354pegase, and Case2848rte). It is shown that the proposed GNN surrogates are 250 to 800 times faster than numerical solvers at predicting the grid state, and they enable fast as well as accurate risk quantification for power grids. It is also shown that that directly predicting aggregated zonal/system level quantities leads to more accurate predictions than aggregating bus level predictions. • Graph neural network proxies for efficient DC OPF computation • Explicit, sampling-based power grid operational risk quantification • Multiple failure modes at system, zone and transmission line levels • Numerical experiments on small to large sized benchmark grids