Study On Interpretability Of Machine Learning Models For Power System Stability Evaluation

Stability is the core concern of power system operation.Machine learning is a promising technology to solve stability assessment problems in the new generation power system.Many researchers have proved its high efficiency and outstanding accuracy.However,most machine learning algorithms build "black box" models which are lack of interpretability.The interpretability of the stability assessment machine learning model can help operators trust the assessment results.As well as provide a reference for stability control.Therefore,it helps eliminate concerns about model reliability and practicality.To enhance the interpretability of the model,this thesis proposed interpretation algorithms for stability evaluation machine learning models.The study follows the idea from local interpretation to global interpretation.First,for a single data and its stability evaluation result,the sensitivity and contribution of features are calculated to reveal the inner logic between the system state variables and stability level.Then a state data clustering method is proposed,and the cluster interpretation is constructed.Finally,by calculating feature importance and partial dependence function,global interpretation is given from the feature perspective.The main work is summarized as follows:A model agnostic interpretation method for calculating the feature sensitivity is proposed based on linear surrogate model.The local linear surrogate model is constructed in the neighborhood of interpreted data based on weighted linear regression and regularization.The sensitivities of features are represented by the surrogate model’s parameters.There are high correlations between power system state variables.Therefore,elastic net is used for feature selection,which makes the interpretation easier to understand.In addition,a training data sampling method considering feature correlation is proposed to make the sampled data more consistent with the original data distribution.The simulations verify the accuracy of sensitivity interpretation.Results show that the proposed data sampling and feature selection methods improve the precision of the surrogate model.Three model agnostic calculation methods for calculating the feature contribution is proposed.First,the Shapley value in cooperative game theory is used to represent feature contributions in a stability evaluation model.The Shapley value calculation method in a machine learning model is given.Then,to solve the calculation difficulty of Shapley value when the number of features is large.An efficient Shapley value calculation method is proposed.By constructing kernel function and simplifying the form of training data,the parameters of the contribution surrogate model are equal to the features’ Shapley values.Finally,because the aforementioned methods take a long time calculating the label of the training data.A feature contribution evaluation method is proposed based on the idea of local linear fitting.The method construction function that maps simplified data to the original space to quickly calculate training labels.The simulations compare the time complexity of the three methods,verify the accuracy of the contribution interpretation,and analyze their application scenarios.In order to get more general interpretations,a state data clustering method is proposed.The cluster interpretation is obtained by surrogate models of representative data.First,the outliers of state data are selected based on isolation forests,and the principal component analysis method is used for dimension reduction of non-outlier data.Then,the Gaussian mixture model(GMM)is used to cluster the processed data.The variational inference is used to select the best number of clusters when solving the GMM.The interpretation index of a cluster is constructed based on the sensitivity and contribution of its representative data.The simulation results show the superiority of the proposed clustering method.Besides,the correctness of the cluster interpretation results is verified.Feature importance calculation methods and partial dependence function calculation methods are proposed to get a global interpretation.The results of existing feature importance calculation methods may have biases.Therefore,a method based on hierarchical clustering is proposed to solve this problem.Two methods are proposed to calculate partial dependence function,in which the accumulated local effects method can be applied to highly correlated features.The simulation results show the bias in existing feature importance calculation methods.And the effectiveness of the hierarchical clustering method and partial dependence functions are verified.