Transient Stability Assessment Of Power System Based On Deep Belief Network

With the rapid development of interconnected power grid in China,the traditional transient stability analysis method based on the model is gradually difficult to meet the needs of online application,while massive data and artificial intelligence technology provide a new approach for the online transient stability assessment(TSA)of power system.Supported by the National Key R&D Program(2018YFB0904500),this thesis deeply studies a transient stability assessment method of power system based on deep belief network(DBN).The main research contents are as follows:(1)In order to study the influence of input features on TSA and find the feature subspace of the ensemble model,different input feature sets are constructed.From the original features including power angle and output power of generator,bus voltage and line power flow,the feature set selected by mutual information filter,the statistical feature set based on power angle trajectory cluster and the feature set of high hidden layer based on stacked denoising auto-encoder are extracted respectively.The three input feature sets can fully reflect the operating state of power system and enhance the separability of the original data,which lays the foundation for the establishment of subsequent single model and ensemble model.(2)In order to achieve fast and accurate transient stability assessment,a TSA model based on single DBN is established.Firstly,according to the characteristics of TSA,this thesis optimizes the DBN algorithm and proposes an experimental method to determine the structure of DBN.Then,using DBN to mine the mapping relationship between input features and transient stability,a TSA model based on DBN is obtained,which has high prediction accuracy,strong robustness and fast prediction speed.Its visualization results of each layer directly show the strong capability of feature extraction of DBN.Finally,it explores the influence of DBN structure and input features on TSA model.It is found that the optimal structure of DBN is not unique,and different input features have their own merits,thus leading to the idea of ensemble model.(3)In order to further reduce the judgment of false positives and false negatives,and achieve reliable and real-time assessment,a two-stage transient stability assessment method based on ensemble DBN is proposed.Firstly,training the DBN models with different structures under different input feature sets,and establishing the ensemble DBN model according to the average mechanism of probability.This ensemble mode of combining multiple input features effectively improves the prediction performance,robustness and generalization ability of TSA.Then,the regression prediction models of stability degree and instability degree based on DBN are established.Finally,it proposes a two-stage assessment process in time sequence according to the credibility.In the first stage,the ensemble DBN model is used to judge the transient stability of power system.If the result is credible,it will enter the second stage to further predict the transient stability margin,otherwise the judgment will continue over time.As a result,it not only greatly reduces the misjudgment and takes into account the rapidity and accuracy of the assessment,but also can accurately measure the stability degree or instability degree of power system after fault,providing more abundant reference information.(4)In order to realize the self-adaptive assessment of transient stability to track the changes of system operating conditions,this thesis proposes an adaptive model updating method based on transfer learning.Firstly,according to the adaptability of the original model after the system changes,two transfer learning schemes,local fine-tuning or overall fine-tuning of the model,can be selected.Then,a method of actively selecting key samples by DBN is proposed to reduce training samples and accelerate the speed of model updating.Finally,the simulation result shows that the proposed method can flexibly restore and improve the performance of the original model in a short time,and achieve the self-adaptive assessment of transient stability.