基于深度强化学习的新能源配电网双时间尺度无功电压优化

Optimization of Dual-time Scale Reactive Voltage for Distribution Network With Renewable Energy Based on Deep Reinforcement Learning

新能源大量接入配电网,其波动性及间歇性容易导致配电网电压的频繁波动问题。传统基于模型的无功电压优化方法高度依托于电网的精准建模,其求解精度与计算速度难以满足含新能源配电网对于电压控制的要求。该文基于深度强化学习,提出一种双时间尺度配电网无功电压优化方法。该方法将电力系统无功电压优化问题转化为马尔可夫决策过程,统筹考虑无功补偿设备的差异化调节特性和不同深度强化学习算法的特点,设计针对离散型设备和连续型设备协调控制的双时间尺度优化方案。其中,长时间尺度上制定并联电容器组投切计划,以调整电压偏移,同时最小化全系统网损;短时间尺度上设置滚动预测窗,制定SVG出力计划,以跟踪电压变化,解决新能源并网带来的配电网电压频繁波动问题。最后通过IEEE33节点拓展系统验证该数据驱动方案在无功电压优化的实现速度和效果上所具有的优势。

A large amount of renewable energy is connected to the distribution network,and its volatility and intermittency are easy to cause frequent voltage fluctuation in distribution network.Traditional model-based reactive power and voltage optimization methods are highly dependent on the accurate modeling of the power grid,and their solution accuracy and calculation speed are difficult to meet the requirements for voltage control of distribution network with renewable energy.In this paper,a dual-time scale distribution network reactive power and voltage optimization method is proposed based on deep reinforcement learning.This method transforms the reactive voltage optimization problem of the power system into a Markov decision process,considers the differential regulation characteristics of reactive power compensation equipment and the characteristics of different deep reinforcement learning algorithms,and designs a dual-time scale optimization scheme by coordinated controlling for discrete equipment and continuous equipment.Among them,the switching plan of the shunt capacitor bank is formulated on the long-term scale to adjust the voltage deviation and minimize the network loss of the whole system;on the short-time scale,a rolling prediction window is set,and the SVG output plan is formulated to track the voltage change and solve the problem of frequent voltage fluctuation of distribution network due to the connection of renewable energy.Finally,the advantages of the data-driven scheme in the realization speed and effect of reactive power and voltage optimization are verified by an IEEE-33 node extension system.