Machine learning based binding contingency pre-selection for AC-PSCOPF calculations

We propose to use off-line machine learning to train an oracle predicting the set of binding contingencies for an Alternating Current Preventive Security-Constrained Optimal Power Flow (AC-PSCOPF) solver. On-line, the oracle's predictions are used instead of the full set of all postulated contingencies, as an input to the PSCOPF solver. A Steady-State Security Assessment (SSSA) is applied to the resulting PSCOPF solution to check the absence of false negatives. Our oracle is a deep neural network multi-label classifier that uses as inputs active and reactive loads, generations, and power flows, computed by an OPF using the same cost function and base-case constraints as the PSCOPF. The proposal is show-cased on the Nordic32 benchmark.