Machine learning-driven multi-agent-based AC optimal power flow with effective dataset creation for data privacy and interoperability

Abstract

As power systems continue to evolve, the demand for effective collaboration among institutions has grown, driven by the challenges of balancing production and consumption, as well as by the increasing need for redispatch. Despite this, achieving interoperability in such a complex landscape is often hindered by concerns regarding data privacy. In response to these challenges, our paper presents a novel approach: a multi-agent system (MAS)-based AC optimal power flow (AC-OPF), empowered by machine learning (ML), designed for safeguarding data privacy and promoting interoperability. In the proposed method, the technical operator agent creates an effective dataset using n-ball, multivariate Gaussian distribution (MGD), and perturbation techniques. It also formulates valid inequalities to reduce the search space. Then, neural network (NN) models are developed to map the feasible space of the AC-OPF by utilizing only active power. Notably, these models conceal both the grid topology and sensitive data before transmission to another agent. Subsequently, the market operator agent integrates these NN models and valid inequalities into a mixed-integer linear programming (MILP) problem. This resulting MILP can be solved with various market based objective functions and constraints considering the power system limits. Thus, if there are private market-based data, they are kept confidential without being shared with the other agent. In addition, mapping is performed using the effective dataset generation technique that ensures a balanced representation of feasible and infeasible data points around the boundary. Furthermore, this effective dataset contributes to achieving remarkable accuracy in NN models, even with a relatively small volume of data points. The results on 30-, 57-, and 162-bus benchmark models of PGLib-OPF demonstrate that the proposed method can be effectively conducted while concurrently enhancing data privacy, and thus interoperability among institutions.