基于GA优化GRU-LSTM-FC组合网络的风电场动态等值建模

Wind farm dynamic equivalent modeling by GA-optimized GRU-LSTM-FC combined network

针对风电场动态等值建模依赖于运行方式和特定扰动,难以获得普适性强的通用等值模型的难题,提出了基于门控循环单元-长短期记忆-全连接(GRU-LSTM-FC)组合网络的数据驱动建模方法,并提出基于遗传算法(GA)对组合网络模型进行调优。首先将风电机组描述为一组微分代数方程组,模型输入为测风塔风速、风向和公共耦合点处的电压时间序列,模型输出为风电场功率时间序列。然后对比了具有记忆作用的LSTM(GRU)网络结构与风电机组微分方程的相似性,以及FC网络结构与风电机组代数方程的相似性,提出基于GRU-LSTM-FC组合网络的风电场等值建模方法。为对组合网络进行模型调优,利用GA优化组合网络中的FC层数和各层神经元数目。最后以某风电场为例验证了所提组合网络进行风电场等值建模的可行性,并将所提方法与其他神经网络模型进行了对比,分析了所提模型的优越性。

Aiming at the dynamic equivalent modeling of a wind farm depends on its operation mode and specific disturbance,which is difficult to obtain a general equivalent model,a data-driven modeling method based on gate recurrent unit-long short term memory-full connection(GRU-LSTM-FC)combined network is proposed,and the genetic algorithm(GA)-based method is proposed to optimize the combined network model.Firstly,the wind turbine is described as a set of differential-algebraic equations.The model input is the wind speed,wind direction at the anemometer tower,and the time series of voltage at the point of common coupling(PCC),and the model output is the time series of the wind farm power.Then,by comparing the similarity between the LSTM/GRU network structure with memory ability and the differential equation of wind turbine,and the similarity between the FC network structure and the algebraic equation of wind turbine,an equivalent modeling method of wind farm based on GRU-LSTM-FC combined network is proposed.In order to optimize the combined model,GA is applied to optimize the number of FC layers and the number of neurons at each layer in the combined network.Finally,taking a wind farm as an example,the feasibility of data-driven equivalent modeling with the proposed combined network is verified,and the proposed method is compared with other neural network models,and the advantages of the proposed model are analyzed.