The Application Of Machine Learning And Network Embedding Algorithms In Power System Transient Stability And Voltage Stability Assessment

Recent years,the power system is approaching its stability margin gradually because of the rapid economic and social development and the rapidly increasing loads.How to ensure that the power system does not suffer from serious instability problems and avoid the blackouts has become an important issue.Power system transient stability and voltage stability are important aspects of power system stability problems.Power system transient stability studies whether the system can transit to a new stable operation state without instability after a severe disturbance.Power system voltage stability mainly studies the margin between the current operation mode of the system and the voltage collapse point of the system.The traditional solution of both kinds of problems is simulation calculation.By giving the parameters of each part of the power network,the corresponding physical process is simulated and the numerical solution is given.The traditional modeling method of power system transient stability problem is to build differential-algebraic equations for each component and the power network in the system,and then use various numerical methods to calculate the post-fault results in time domain.The problem of voltage stability usually adopts the method of continuous power flow calculation.Starting from the current operation mode of the system,the load of the system is gradually increased until the voltage collapse point of the system is obtained.Two kinds of simulation methods are both time consuming,and are difficult to meet the requirements of online using.In this paper,machine learning and network embedding methods are used to study the power system transient stability problem and voltage stability problem.The research contents and the results are as follows:(1)For the problem of power system transient stability,a set of electrical features reflecting the steady-state of the power system is proposed,and the problem of transient stability assessment is categorized into the classification problem in machine learning.The XGBoost algorithm,a machine learning frontier algorithm,is used to assess the transient stability and it is modified according to the background of transient stability problem.The concept of transient stability loss function is proposed,and the output of the model is probabilized to evaluate the reliability of the prediction of the model.For the data acquisition,the method of accumulating the samples of different operation modes is studied.The case study results show that the modified XGBoost algorithm is superior to other common machine learning methods in accuracy and recall rate.The definition of the transient stability loss function makes the model less prone to mis-classification problem in prediction,it can avoid operator's neglect of serious instability.The probabilistic output enables that the reliability of the model prediction can be measured,so that some wrong predictions can be avoided.(2)The main problem of the traditional machine learning based transient stability assessment method for the power system is that,the trained model can only be used to predict the consequences of the faults at specific locations,and its generality is poor.To solve this problem,the network embedding algorithm,which is the latest achievement in the field of artificial intelligence,is used to extract the features of the power network nodes.The features of power network nodes can be obtained by combining the information of structure and electrical quantity.Thus,a general model can be trained to evaluate the fault consequences of any node in the power network.According to the background of power system transient stability problem,the application of the TADW algorithm in network embedding is studied.Then,the Support vector regression is used to model the relationship between the TADW features and the fault consequence of the nodes.For the data acquisition,the method of accumulating node failure consequence samples for different operation modes is studied.The case study results show that the features of the power network nodes extracted by TADW algorithm can reflect the fault consequences of the nodes under specific operation modes.The model trained by the sample set of such features can evaluate the transient stability consequences of any node under any operation mode,thus it can meet the demand of generality.At the same time,the consequences of power network nodes failure given by the model are used to rank the importance of the nodes.Compared with the traditional index based ranking methods,it can give a more reasonable result.(3)To solve the power system voltage stability problem,the XGBoost regression model is established by using the electrical feature set reflecting the steady state of power system.The case study results show that,the XGBoost model has better performance than other machine learning methods in the aspect of accuracy and calculation speed.Then,the fault tolerance performance of the model is investigated in view of the possible errors in real-time operation information measured by PMU,the results show that when the two common measurement errors occur,XGBoost model can give relatively reasonable prediction results.Because the machine learning model will inevitably lead to some false predictions,this paper studies the problem of model updating with the samples which have a large error.The results show that using such samples to update the model iteratively can make the model no longer suffer from similar errors in subsequent use.