Research On Fast Assessment Of Transient Voltage Stability Of AC/DC Receiving-End Power Grid

With the intensive commissioning of high voltage direct current projects,a large number of new energy access and load composition and characteristics changes,the proportion of fast dynamic response components is increasing,the problem of voltage stability is prominent,the coupling between AC/DC and sending/receiving-end is becoming tighter,and the power outage increased risk.The traditional numerical simulation method based on detailed mathematical model has a slow calculation speed and it is difficult to meet the requirements of online transient voltage stability assessment.The methods based on data mining and machine learning provide new ideas for solving this problem.This thesis is based on data mining and deep learning technology,the main research work and results achieved are as follows:(1)The main factors affecting the transient voltage stability of AC/DC receiving-end power grids is studied.It includes 4 aspects:impact of transmission network capacity limitation,impact of induction motor load,Impact of reactive power support capacity constraints of receiving-end power grid and interaction influence of transient rotor angle problem.The discussion of factors affecting the transient voltage stability provides the basis for the selection of characteristics of transient voltage stability assessment based on machine learning.(2)A method of partitioning AC/DC power grids based on t-stochastic neighbor embedding(t-SNE)and fuzzy C-mean(FCM)algorithms is developed.The grid partitioning is conducive to the selection of key buses for rapid assessment and can reflect the overall transient voltage stability level of each region.For regional power grids,the number of buses is small.A transient voltage sag area matrix is constructed based on the transient voltage time series information,which is mapped to a two-dimensional plane using the t-SNE algorithm,and the receiving-end grid is partitioned to allow for the characteristic of fast dynamic response components;for a large-scale power grid with many buses and branches,the transient voltage sag area matrix is hard to constructed,so the electrical distance matrix is constructed based on steady-state network information.First,the electrical distance matrix is reduced to two dimensions using the t-SNE algorithm,and then FCM clustering is performed,and Xie-Beni coefficient is used as an evaluation index to produces the partition results.A Shandong 500kV and above main grid system and a east china power grid were used for verification.The verification results show that the proposed partitioning method is reasonable and effective.(3)A method for fast assessment of transient voltage stability of AC-DC receiving-end power grids based on convolutional neural network(CNN)is established.According to the relative distance of the buses,the steady-state power flow characteristics of each zone are selected to construct the line fault severity index.The line number of the fault is coded according to the code,and the coding result and the line number of the fault are used as the fault characteristics to improve adaptability to different faults.Using a one-dimensional convolutional neural network to build a transient voltage stability security risk situation awareness model can reduce the loss of information and make full use of the ability of CNN’s local perception.The particle swarm optimization technology is used to determine the optimal convolution kernel size and number of each partition to improve the performance of CNN.A transient voltage stability assessment module for large-scale power grid security risk situational awareness early warning demonstration software was written based on Python,which can implement functions such as online application and online update.The proposed method is verified by the Shandong 500kV and above main grid system.The analysis results of examples show that the proposed CNN-based transient voltage stability assessment method has higher assessment accuracy than other models;it can reduce the accuracy without reducing the accuracy.The redundancy of features is decreased and the speed of online application is improved.