Research On Harmonic State Estimation Based On Improved GAN

In recent years,with the continuous development of society,economy and science and technology in China,the level of electrical automation is constantly improving,the number of nonlinear power and electronic equipment in the power system continues to grow,and the power quality problems directly caused by harmonics continue to increase.Therefore,in order to ensure the safe and effective operation of the power grid,it is very important to accurately estimate the harmonic state in the power grid to effectively control the harmonic pollution.Traditional harmonic state estimation adopts the mechanism analysis method based on physical model,describes the coupling relationship between nodes,and estimates the harmonic state of unmonitored nodes through the known monitoring data,including harmonic voltage and harmonic current,etc.However,due to the limitation of few measuring devices,the difficulty in obtaining accurate harmonic impedance,the complexity of network topology and the variation of power grid operation mode,the measurement equation is undetermined,the system is not globally observable and the coupling relationship between nodes is difficult to be accurately extracted.Depth compared with other traditional machine learning methods,learning more widely applicable to high dimension,a large quantity of data application scenario,its characteristic is using the data of batch training,training through iterative function to minimize the loss,repeatedly adjust the depth of the neural network,the weight of each hidden layer unit parameters as input and output height to fit with the sample label,The mapping relationship between them will be solidified as the weight parameter of hidden layer.The data-driven relationship extraction method avoids the disadvantages of physical model,and can reflect the coupling relationship between nodes under the influence of uncertain factors such as high-dimensional,nonlinear and strong time-varying.Generating adversarial network avoids the difficulty of loss function design due to the addition of discriminant network.Compared with other deep learning algorithms,generating adversarial network can generate clearer and more real target data.Based on the traditional generative adversation network,a two-node harmonic state estimation method based on pix2pix and a multi-node harmonic state estimation method based on the fusion refining network are proposed in this paper.Based on the batch real sample data of node pairs,the model extracted the coupling relationship between node pairs through the conditional generation adversation game,and generated the harmonic data of target nodes in line with the characteristics of the measurement data of distribution network.Finally,on the basis of solidifying the coupling relationship between node pairs,the state data of target node is accurately estimated based on the monitoring node data.In this paper,an IEEE 14-node simulation model was built with PSCAD for verification.The experimental results show that the proposed method can effectively estimate the harmonic current and harmonic voltage of the target node,and realize the data-driven harmonic state estimation method.The anti-noise test also verifies the anti-interference and generalization ability of the proposed method.