Research On Frequency Nadir Prediction Of Post-disturbance Power Grid With New Energy Based On Deep Learning

Challenged by large-scale integration of new energies,the frequency stability of modern power system has attracted great attention.The frequency nadir prediction of power grid after disturbances is crucial to online frequency stability assessment.To predict fast and accurately provides a valuable reference for frequency emergency control strategies.The dynamic frequency is a complex variable related to space and time.To make full use of the spatiotemporal correlation involved in measurement data,and give consideration to both calculation speed and accuracy,this paper conducts related research on dynamic frequency prediction of power system based on convolutional long short-term memory(ConvLSTM)network.The main contents of this paper are summarized as follows:(1)Aiming at the characteristics of power system dynamic frequency,a method for predicting the frequency nadir of post-disturbance power grid is proposed.The proposed method is based on ConvLSTM which combines the advantages of long short-term memory(LSTM)and convolutional neural network(CNN).The analysis of principles shows the superiority of this method in power system dynamic frequency prediction.(2)Aiming at center of inertia(COI)frequency prediction,the detailed modeling process of ConvLSTM prediction model is given,including the selection of input features,sample organization and generation based on PSS/E,and the construction of input tensors.On the New England system,the impacts of model structure,the length of input time series,and distribution plane of electrical nodes on the predictive model are analyzed.Compared with the other machine learning methods(e-support vector regression,1D-CNN and LSTM),the proposed method shows better performance.To study the impact of incomplete measurement data on the ConvLSTM-based prediction model,two cases are designed from the perspective of data loss and incomplete configuration of measurement devices.The results show that this prediction model has strong robustness and learning ability.Finally,the proposed method is verified in an actual power system of South Carolina,and the results shows that the proposed prediction method is possible to apply in large-scale systems.(3)In order to better construct the sample database of post-disturbance power grid containing new energy,the sample space is expanded on the basis of original single disturbance,that is,samples of cascading failures are added.Specifically,a series of disturbances of line fault,line tripping,fault clearing,wind generation tripping,and line reclosing are simulated,and then COI frequency nadir is predicted using the proposed ConvLSTM-based method.Next,considering the temporal and spatial distribution characteristics of dynamic frequency response,generator frequency nadir prediction under single disturbance of machine tripping is realized based on ConvLSTM,in order to formulate more efficient frequency emergency control strategies.The prediction is verified in a modified New England system with wind power.