Power Grid Transient Stability Margin Assessment Based On Power Flow Mapping And Deep Learning

With the continuous expansion of the scale of power grid interconnection,the mechanism model and stability problems of power system are becoming more and more complex.There are several traditional transient stability assessment(TSA)methods such as time-domain simulation,transient energy function and extended equal-area criterion.However,the flaws of these methods such as their poor model adaptability and complexity of computation are salient.Therefore,it is difficult to meet the time efficiency and accuracy requirements of current power grid.Thanks to the wide area measurement system(WAMS),it provides a way to achieve datadriven transient stability analysis.If only the steady-state characteristics are used as input,it is called the TSA under the concept of security region.Because the disturbance has not occurred,if there is instability risk in the system,the operator has enough time to take preventive measures,which has strong practical application value.In recent years,due to its strong self-learning ability,deep learning algorithm has gradually become a powerful tool for data-driven TSA.As one of them,convolutional neural network(CNN)has strong image recognition ability and can effectively extract the spatiotemporal features of the power grid.Therefore,based on power flow mapping and composite CNN,an assessment method of power system transient stability margin is proposed.The main research contents of this paper are as follows:(1)Taking account of the advantages of CNN in image recognition,a power flow map mapping method is proposed based on the real-time measurement data of power grid.With the help of geographic information system(GIS),the visualization of steady-state power flow and topology information is realized.The generated computer-vision-based power flow image(CVPFI)can effectively represent the operation state of power grid from different aspects,which provides a new way for the input characteristics of TSA model.(2)Given the anticipated contingency set,a composite CNN is constructed to quantitatively map the relationship between the steady-state characteristics of power grid and the generator stability margin.After the samples are divided into different categories by improved CNN classification model,the CNN regression model is further used to predict the generator stability margin of various samples.Moreover,the defect of misclassification of critical stable samples is improved by building a correction network,which effectively improves the prediction accuracy.(3)A general interpretation method of deep learning TSA model is proposed.Based on weighted linear regression and regularization,a local linear surrogate model is constructed to approximate CNN regression model in small neighborhood,which reveals the operation logic of deep learning model to a certain extent.The interpretation results can help dispatchers to establish confidence in the deep learning model.The simulation results of IEEE-39 bus system verify the effectiveness of the proposed method.It can quickly estimate the transient stability margin under the anticipated contingency only through the steady-state information.Moreover,it has great generalization ability and high prediction accuracy,and can provide important reference for dispatchers to take control measures.