Dynamic Distribution Network Reconfiguration With Generation and Load Uncertainty
Shahab Bahrami, Yu Christine Chen, Vincent W. S. Wong
Abstract
Given the uncertainty in load demand and renewable energy sources, the distribution network reconfiguration (DNR) problem is a stochastic mixed-integer nonlinear optimization program with a running time that scales exponentially with the number of sectional and tie line switches. Stochastic optimization techniques require knowledge of the stochastic processes of the uncertain parameters, which may not be available in practice. This paper addresses both the scalability and uncertainty issues in solving the DNR problem by developing a deep reinforcement learning (DRL) algorithm that determines the optimal topology using a transformer deep neural network (DNN) architecture, and subsequently solves an AC optimal power flow (OPF) problem to satisfy the operation constraints. A neural combinatorial optimization algorithm is applied to train the DNN, which penalizes infeasible solutions. Simulations on a 119-bus test system show that our proposed algorithm can obtain a near-optimal solution to the stochastic DNR problem with a small gap (i.e., 4.7% on average) from the objective value of the deterministic DNR problem. When compared with existing learning-based DNR algorithms in the literature, our proposed algorithm can obtain at least 11% lower objective value. We demonstrate the scalability of our proposed algorithm in larger systems with 595, 1190, and 3570 buses.