A Method For Power System Dynamic Security Assessment Based On Image Recognition

As the scale of power system in China grows larger and its structure becomes more complex,the task of power system dynamic security assessment(DSA)is becoming more and more onerous.Screening a huge amount of contingencies and classifying the stable and unstable faults sets are important works for DSA.The developing of machine learning technology has provided a possible way to rapidly distinguish the power system transient stability result.This paper mainly discusses the application of image recognition for power system DSA,the main work is given as follows:Firstly,a power system data samples generating platform has been developed based on Python tkinter module,which can compile the PSASP data file and generate a lot of calculation samples automatically by calling the transient stability program.The samples can be used in the training process of many intelligent algorithms.Secondly,a method for transient stability discrimination based on image recognition is proposed in this paper.At first,a short time time-domain simulation is performed to get the physical quantities such as the rotor angle,frequency and weighting accelerating power of each generator.Then a convolution neural network(CNN)is utilized to recognize the changing curves of these physical quantities and distinguish the transient stability results.Tests are carried out in the New England 39-bus system and the comparison result of these physical quantities has been obtained.Thirdly,a method for transient stability contingency screening is studied in this paper.Based on the output probability in softmax layer under different simulation durations,the corresponding CNN can screen the stable and unstable samples layer by layer.The method is verified in the IEEE 39-bus system and tests are also carried out in the IEEE 300-bus system and the provicial 433-bus system.The improved screening method for complex fault sets which have many multi-swing instability and critical instability samples is discussed.Since the training of CNN can be performed off-line and the model is suitable for on-line analysis,the proposed contingency screening method is valuable for practical application.