A Reinforcement Learning Based Voltage Regulation Strategy for Active Distribution Networks

The penetration of renewable energy in active distribution networks is gradually increasing. Conventional analytical-based numerical solution methods face a serious challenge with the need for faster response times for intermittent PV power generation. In this paper, a recurrent neural network strategy is proposed to optimize the reactive power value of a PV inverter to achieve a fast solution to the voltage overrun problem. This model-free fast optimization strategy formulates the reactive power optimization problem as an observable Markov decision process. The recurrent neural network in the hidden layer of the model-free strategy successfully integrates time series information into the continuous control of the distribution network. A case study of a typical 33-bus system validates the effectiveness and good performance of the method.

I. Introduction

More and more solar power sources are being installed in the distribution network, and the intermittent and rapid fluctuations of PV power generation will lead to serious voltage overrun problems [1]. In response to the voltage limit crossing caused by the continuously increasing PV penetration, the IEEE 1547 standard allows smart PV inverters to dynamically supply reactive power to participate in the regulation of the distribution network voltage [2]. Coordinated control of widely distributed PV reactive power output is formulated as an optimization problem[3]. It also gives rise to many flexible optimization schemes, such as [4] –[6]. However, the above optimization problem requires repeated iterative operations by specific solving tools to converge to the optimal solution. Considering the rapid fluctuation characteristics of PV power generation, the optimal results obtained from the above scheme may deviate from the actual PV power generation in practical applications [7].