Power System Transient Stability Assessment Based On Deep Learning

In the future,in order to improve its own efficiency and economic value,the operation of the power system may be closer to the stability limit,a large number of new energy access,as well as the power marketization reform makes the system dynamic changes more complex,the traditional model-based power system stability assessment methods have been difficult to meet the requirements of the online dispatch of the grid.At the same time,the number of Wide Area Measurement Systems(WAMS)and Phasor Measurement Units(PMU)placed in the grid has been increasing,effectively improving the situational awareness of the system,coupled with the continuous improvement of computer computing power,the data-driven transient stability assessment(TSA)method has been widely studied.However,TSA based on traditional machine learning methods has the problem of complex feature construction and fixed model structure.In this thesis,a transient stability online assessment system based on a Recurrent Neural Network(RNN)is proposed,which has an assessment time adaptive function and can weigh the relationship between assessment accuracy and assessment speed to output transient stability assessment(TSA)results earlier.The specific research contents are as follows:Firstly,this paper proposes two TSA feature sets based on PMU data characteristics and deep learning method advantages.Following the feature set this paper experiments in IEEE-39 system using time-domain simulation,constructs labeled samples of TSA under different operation modes,and proposes performance evaluation metrics for the unbalanced sample model.Secondly,two special structures of RNN are selected to build TSA models based on Long Short-Term Memory(LSTM)network and Gated Recurrent Unit(GRU),which are trained and tested in IEEE-39 system against the performance of different models.The simulation results show that when line power is used as an assessment feature,the GRU-based TSA has the highest accuracy and the best generalization,and the RNN-based TSA has a significant improvement in accuracy and generalization compared to the traditional machine learning method.Finally,the GRU network model is refined to construct an online TSA system that assessment time adaptation.The system uses PMU as the system data input,which can judge the reliability of the TSA at all times of the system and output the assessment results when the assessment is found to be reliable.The assessment system was trained and tested in IEEE-39 system experiments to determine the size of the parameters at optimal performance of the system,and the experimental results showed that the average assessment response time of the system at optimal performance was 2.15 cycles,and the assessment accuracy was 99.85%.