Generative Design for Power System Networks Using WGAN and Graph Performance Measures for Guided Generation

Abstract Interdependent critical infrastructures are prone to be a target of disruptions from either natural disasters or deliberate attacks that can cause a great magnitude of negative consequences on the system. To withstand and reduce the impacts of disruptive events, it is beneficial to consider them in the early stage of design to create a configuration that can maintain its performance as much as possible. However, due to the complexity of modern systems, it is difficult to directly optimize the system design of large-scale with complicated constraints. Hence, in this research, a generative approach based on Wasserstein GAN is implemented to create intelligent designs. The power system networks are represented as graphs with features holding information on the nodes and edges that relate to electrical components. The trained generator can create a power system network consisting of generators and loads, where the buses are connected to transmission lines of a certain capacity. Afterwards, performance measures are employed to induce the generation of a population with more appealing characteristics. The framework was applied to a case study of a 57-bus system of synthesized power grid and the results have shown that the model was able to generate feasible designs of higher resilience.