Power System Transient Security Analysis And Assess Based On Interpretable Machine Learning Model

With the proposal of the goal of "carbon peak,carbon neutral" and the promotion of a high proportion of new energy grid connection,the proportion of photovoltaic power generation and wind power generation in the total installed capacity continues to increase,and more and more UHV AC/DC lines are put into operation,exacerbating the complexity of the power grid and making the issue of security and stability even more prominent.How to quickly and accurately analyze the transient security of power systems is the key to maintaining the safe and stable operation of new power systems.The transient security analysis method based on machine learning can extract security and stability information from a large number of features,but it is generally weak in interpretability.This paper focuses on how to select key features from power grid characteristics,build a machine learning model for transient security assessment,and improve the interpretability and practicality of assessment and analysis.The research content is as follows:In terms of power grid transient security feature selection,this paper proposes two improved feature selection methods to address the high complexity of the system and the inability to fully reflect the transient security characteristics of the system solely by relying on key power flow sections,in order to reflect the most comprehensive system characteristics with as few feature subsets as possible.One is that when constructing the original power system transient security feature set,it not only contains the power flow information of the system,but also adds feature indicators based on expert domain knowledge,known as"expert features."."Expert features" are used as known features for feature selection."Expert characteristics" reflect the physical mapping between expert domain knowledge and power system transient security characteristics,which can improve the interpretability of feature selection results.The second is to improve the criteria for evaluating the importance of features,and propose a weighted conditional mutual information based feature selection method(WCMI),which can eliminate redundant features,simplify feature subsets,and obtain key features for power system transient security analysis while ensuring the accuracy of model prediction.In the field of power grid transient security assessment,in order to solve the problem of avoiding false positives and causing hidden dangers of missed alarms in transient security assessment,an improved support vector machine algorithm that considers the rate of missed alarms and the rate of false positives is proposed,which are respectively referred to as radical support vector machine(ASVM)and conservative support vector machine(CSVM).This method defines a gray zone where stable and unstable regions intersect,and performs security evaluation through CSVM and ASVM to ensure that the evaluation results of samples outside the gray zone are accurate and credible,without missing or false alarms.In terms of the interpretation of transient security assessment models,this paper proposes a model interpretation method based on SHAP(Shapley Additive exPLANATIONS)analysis,which can better help us understand the transient security analysis model,in response to the widespread problem of poor interpretability of machine learning models.The model is interpreted globally and locally through SHAP calculations.Through global interpretation,feature importance ranking and score weights are obtained,further verifying the effectiveness of the WCMI algorithm proposed in this paper;Through local interpretation,the contribution of each key feature value of an individual sample to the transient security assessment results is obtained.For the samples in the gray zone,this article uses the SHAP local interpretation method to analyze them,and combines this part of the samples with key features in the local analysis to conduct model training,further reducing the scope of the gray zone.This thesis focuses on the bottleneck issues in power system transient security analysis based on machine learning models.Through improvements in feature selection,security analysis algorithms,and model interpretation,the accuracy and interpretability of power system transient security analysis models have been improved.