Static Voltage Stability Limit Sample Generation Based On Transfer Learning And Generation Adversarial Network

The continuous expansion of the scale of the power grid,the increase in the penetration rate of renewable energy,and the emergence of electric vehicles have introduced more temporal and spatial uncertainty and randomness,as well as more data,to the operation of the power system.The operation of power systems is becoming more and more complex,and data-driven AI methods will be an effective way to solve this problem.The key to applying AI methods in the field of power system operation is to obtain sufficient data corresponding to the task,that is,the power flow sample which meets the demand.Although many power flow samples naturally exist in the field of power system operation,these samples are often within the stable region,and there are very few abnormal samples,and it is difficult to deal with tasks related to stable operation,critical or abnormal samples are more needed.Taking static voltage stabilization as an example,the key lies in the generation of static voltage stability limit samples.The traditional method is to generate enough static voltage stability limit samples through multiple point-by-point simulations such as CPF or Po C.However,under the new situation of the power grid,as the system dimension becomes higher,the simulation scale expands,and the number of calls increases sharply,resulting in too long offline calculation time.It may also face the ill-conditioned and non-convergent problems as traditional power flow calculation,which limits the application of AI methods.A new generation method is necessary.This paper proposes a generative model based on transfer learning and GAN,which is used for the generation of static voltage stability limit samples.Firstly,it is noticed that the critical sample is a special kind of power flow sample.Setting the non-contact node voltage as the features of the critical sample helps solve the non-convergence problem of power flow for samples and non-zero injected power for the contact nodes.Secondly,choose the Wasserstein distance as the loss function.At the same time,the spherical uniform distribution is used as the power growth direction and the input of the generation of the GAN,and the superiority of the Wasserstein distance and the spherical uniform distribution is verified on the IEEE14 system.Finally,the WGAN-GP combining transfer learning is established to learn the special constraints and distributions of limit samples,and the minimum singular value of samples is selected to show the quality of samples.The results of the case study on the IEEE118 system show that the combination of WGAN-GP and Transfer Learning can generate high quality samples more effectively than WGAN-GP only.