Deep Reinforcement Learning Based Volt-VAR Optimization in Smart Distribution Systems

This paper develops a model-free volt-VAR optimization (VVO) algorithm via multi-agent deep reinforcement learning (DRL) in unbalanced distribution systems. method is novel since we cast the VVO problem networks to an intelligent Q-network (DQN) framework, which avoids solving specific model directly when facing time-varying operating conditions We consider statuses/ratios of switchable capacitors, voltage regulators, and smart inverters installed at distributed generators as action variables agents. A delicately designed reward function guides these agents interact with system, direction reinforcing regulation power loss reduction simultaneously. The forward-backward sweep for radial three-phase systems provides accurate flow results within few iterations DRL environment. proposed realizes dual goals VVO. test this on IEEE 13-bus 123-bus Numerical simulations validate excellent performance reduction.