Distributed multi-step Q(λ) learning for Optimal Power Flow of large-scale power grids

This paper presents a novel distributed multi-step Q(λ) learning algorithm (DQ(λ)L) based on multi-agent system for solving large-scale multi-objective OPF problem. It does not require any manipulation to the conventional mathematical Optimal Power Flow (OPF) model. Large-scale power system is first partitioned to subsystems and each subsystem is managed by an agent. Each agent adopts the standard multi-step Q(λ) learning algorithm to pursue its own objectives independently and approaches to the global optimal through cooperation and coordination among agents. The proposed DQ(λ)L has been thoroughly studied and tested on the IEEE 9-bus and 118-bus systems. Case studies demonstrated that DQ(λ)L is a feasible and effective for solving multi-objective OPF problem in large-scale complex power grid.

► Multi-objective Optimal Power Flow. ► Distributed Reinforcement Learning dealing with complex multi-objective problem. ► Testing on the IEEE 9-busbar and the IEEE 118-busbar system. ► DQ(λ)L owns fast convergence speed without loss of high convergence precision.