A Representative Training Dataset Generation Method for Neuron Network based Optimal Power Flow Solver

Data-driven Neuron Network solvers can solve optimal power flow problems effectively without explicit information of the power system parameters. However, the training quality of Neuron Networks (NNs) significantly depends on the representativeness of the training dataset. Conventional dataset generation methods often fail to sufficiently cover all possible data scenarios, particularly when dealing with optimization problems involving multiple inequality constraints. This paper introduces a novel dataset generation method designed to ensure representativeness for NN training. Specifically, the proposed method begins by sampling extreme load conditions and identifying the set of active constraints in each corresponding optimal solution. From each discovered active set, additional candidate sets are generated by removing one or more constraints to explore further feasible regions. An feasibility check is then applied, discarding any candidate set that does not yield a valid operating point. For every active set that passes this check, a specialized optimization problem is solved to balance minimizing the original cost function and maximizing load. Finally, the unique (load,optimal dispatch) pairs obtained from all feasible sets are collected. This dataset effectively captures a broad range of operating conditions and constraints, making it a more robust and comprehensive training resource for NN-based OPF solvers. Numerical experiments are conducted to validate the efficacy of the proposed method.