Transient Stability Control Of Power System Based On Deep Learning For Real-time Emergency Load Shedding

Along with the fast developing of the electric network,a great deal of new energy has been brought into the electric network,which has raised the safety and stability of the electric network.The rapid and precise emergency control of the transient stability of the electric power system can strengthen the secondary defence,prevent the destruction of the electric system,and even cause the failure of the network.Emergency load shedding is an effective and commonly used emergency control measure for power system transient stability.Solve the fully optimized load shedding model with large amount of calculation,and slow response to the drastic changes of the system such as the increase of renewable energy and more active demand-side behavior.Other methods based on sensitivity will damage the control accuracy and cannot guarantee the global optimization.AI methods have recently been widely recognized for their real-time decision-making ability to cope with system changes.The existing artificial intelligence methods for emergency load shedding are based on shallow learning algorithms,which can lead to load underswitching and overswitching events.However,when the load is cut off,the power system will be exposed to the high risk of instability after control.This instability will spread to cascade events,resulting in higher costs than the excessive cut off event.This paper focuses on emergency transient stability control and proposes a deep learning method for real-time power system emergency load shedding.The main research contents are as follows:(1)A risk aversion learning algorithm with specially designed loss function is proposed to alleviate the cost imbalance in the training of deep learning model.The loss function transforms the constrained emergency load shedding(ELS)optimization into an unconstrained problem with post-control stability adjustment,aiming at improving the control success rate of ELS and reducing the overall control cost.This risk-averse learning algorithm aims to avoid cascading unstable events caused by insufficient load.(2)The quasi-linear relationship between the extended equal-area criterion(EEAC)margin and the minimum ELS quantity is studied in order to estimate the required minimum load shedding according to the EEAC margin.The minimum load shedding can be used as a reliable reference to predict the stability of the system after control without numerical simulation,which is an effective method to check the effectiveness of control.(3)A deep neural network(DNN)is established as a machine learning model,and the deep learning method is applied to ELS prediction for the first time.The proposed ELS method has been tested in New England 39 bus system and Nordic 41 bus system,including a series of comparative studies with classical DNN and shallow learning algorithm.The results verify the improved prediction accuracy of the deep learning algorithm and the reduced overall control cost of risk aversion DNN.