Optimal Design of Neural Network Structure for Power System Frequency Security Constraints

Abstract

Recently, frequency security is challenged by high uncertainty and low inertia in power system with high penetration of Renewable Energy Sources (RES). In the context of Unit Commitment (UC) problems, frequency security constraints represented by neural networks have been developed and embedded into the optimization problem to represent complicated frequency dynamics. However, there are two major disadvantages related to this technique: the risk of overconfident prediction and poor computational efficiency. To handle these disadvantages, novel methodologies are proposed to optimally design the neural network structure, including the use of asymmetric loss function during the training stage and scientifically selecting neural network size and topology. The effectiveness of the proposed methodologies are validated by case study which reveals the improvement of conservativeness and mitigation of computation performance issues.