基于机群划分与深度强化学习的风电场低电压穿越有功/无功功率联合控制策略

Combined Re/Active Power Control for Wind Farm Under Low Voltage Ride Through Based on Wind Turbines Grouping and Deep Reinforcement Learning

在低电压穿越(low voltage ride through,LVRT)过程中,电网企业要求风电场向系统提供无功支撑;在满足所规定的无功输出基础上,利用风电场剩余容量提供有功功率,对保障系统稳定性意义重大。该文提出一种基于机群划分与改进深度确定性策略梯度(deep deterministic policy gradient,DDPG)的风电场LVRT有功/无功功率联合控制方法。首先,将LVRT期间风机的有功/无功控制分为3个阶段,并基于此构建了风机控制模型;其次,根据风机运行特性将其划分至多个机群,在功率分配过程中,对属于同一机群的风机分配相同的控制指令,该步骤大大降低了优化变量个数与优化问题求解难度;然后,提出一种不含评价网络的并行化DDPG(critic-network free based parallel DDPG,CFP-DDPG)深度-强化学习算法框架,确立了基于CFP-DDPG的风电场功率控制框架,设计控制中的状态量、动作量、评价函数、模型训练策略和控制方法;最后,采用我国某实际风电场数据验证方法的有效性,结果表明,机群划分步骤有助于快速准确得到功率分配方案,CFP-DDPG通过改进动作评价方法并引入并行化结构增强了智能体的探索力,有助于取得更优的控制方案。

During low voltage ride through(LVRT),power grid requires wind farm(WF)to provide reactive power to the system.After satisfying the required reactive power output,the remaining capacity of the WF can be used to provide active power to ensure system stability.This paper proposes a combined active and reactive power control method for WF under low voltage ride through based on wind turbines(WTs)aggregation and improves deep deterministic policy gradient(DDPG).First,the active and reactive power control processes of WTs are divided into three stages,based on which the model of the WT control system is established.Then,the WTs are divided into multiple groups according to their attributes.In the process of power distribution,the same control commands are distributed to the WTs belonging to the same group,and this greatly reduces the number of optimization variables and the difficulty of solving optimization problems.Next,a critic-network free based parallel DDPG(CFP-DDPG)deep reinforcement learning method is designed.On this basis,this paper constructs a CFP-DDPG based power distribution framework through defining state,action,evaluation function,model training process and decision-making method.Finally,the effectiveness of the method is verified by using the real WF data in China.The results demonstrate that the WTs aggregation step could help acquire the power distribution quickly and avoid the algorithm falling into the local optimum.And CFP-DDPG could successfully enhance the exploratory ability of the agent through introducing a parallel structure and improving the evaluation method.