Study On Transient Stability Assessment Of AC/DC Hybrid System Based On Deep Learning

Power system transient stability assessment(TSA)is an effective means of security situational awareness and the basis for early warning.The modern power system is a high-dimensional nonlinear system.With the continuous expansion of the ultra-high voltage AC/DC interconnection,the increasing penetration of renewable energy,and the frequent occurrence of natural disasters,the outage risk of the power system is increasing.The traditional model-based security and stability analysis method is difficult to meet the online dispatching operation requirements,and the data mining method has provided new ideas to solve this problem.The research work and results of this paper mainly include:(1)The construction method of the original feature set of power system transient stability is studied.Firstly,based on the transient stability mechanism of power system,the key influencing factors of transient stability risk are analyzed.Combined with the research results of existing literatures,the original feature set is established including five parts:system-level static features before failure,component-level static features of fault points and their adjacent regions,features that represent the influence of wind power and DC,and fault information.The simulation results of 500kV and above main grids in Shandong Province show that The original feature set not only reflects the overall steady state operation of the power system,but also contains key information that affects the transient stability of the power system,and achieves a balance between the number of features and the amount of information of the original feature set.(2)The key feature extraction method for TSA of AC/DC hybrid system based on mutual information theory and deep learning model is studied.The key feature extraction system for TSA is established:In the first stage,the feature is initially screened based on the mutual information theory,and the feature that the transient stability category is obviously unrelated or obviously redundant with other features is eliminated.In the second stage,the stacked denosing autoencoder(SDAE)is used to perform multi-layer abstract expression on the pre-screened features,and the high-order features of the input are obtained for subsequent dynamic security assessment.The simulation results of 500kV and above main grid system in Shandong Province show that the feature pre-screening can effectively reduce the scale of the follow-up search process and improve the accuracy of the follow-up evaluation.The feature extraction method based on SDAE can obtain more distinguishing features than principle component analysis,REFICE and rondom forest methods.(3)The TSA method of AC/DC hybrid system based on ensemble learning is studied.A TSA method based on support vector machine(SVM)ensemble model is proposed.The deep learning model can obtain high-level abstract representation of the underlying measurement data by using its multi-layer structure;the ensemble learning model has the advantage of better classification or prediction performance than the sub-learners.The SVM ensemble model uses SVM as a sub-learner of the ensemble learning model,and uses the level features of SDAE extraction as input to each sub-learner,which combines the advantages of deep learning and ensemble learning.The SVM ensemble classification model is used to evaluate the transient stability of the power system.The credibility analysis is carried out on the evaluation results,and the input space is divided,which balances the rapidity and accuracy of the assessment.The SVM ensemble regression model is used to predict the transient stability margin of the power system.Based on the effect theory and the proposed transient stability margin index,the severity of the operation scenarios is graded,which makes the transient stability assessment result more intuitive.The simulation results of 10-machine 39-bus system and 500kV and above main grid system in Shandong Province show that the transient stability assessment method based on SVM ensemble model has lower missing alarm rate and false alarm rate than other classification methods,and has generalization ability for unknown scenarios.