Research On Correlation Characterization And Causality Identification Methods Among Variables In Data-driven Transient Stability Assessment Of Power System

As the power grid continues to expand in scale,the operational behavior of power systems is becoming increasingly complex,posing greater challenges to their secure and stable operation.Transient stability assessment is a critical component of power system security analysis,and accurate assessment results are crucial for ensuring the safe operation of power systems.In recent years,with advancements in grid measurements and the development of data analysis and processing methods,data-driven approaches for transient stability analysis in power systems have emerged as a research hotspot.Data-driven methods for transient stability assessment extract variable characteristics from the data and establish virtual mapping relationships between variables,thereby avoiding the difficulties associated with physical modeling.Therefore,accurate feature extraction and the representation of relationships among variables form the foundation of data-driven transient stability assessment.However,current data-driven methods for transient stability assessment suffer from the following limitations: 1)the spatiotemporal correlation characteristics among operational variables have not been fully explored and characterized,resulting in weak robustness and limited predictive performance of the models;2)the causal relationship between input and output variables is not clear,making it difficult to be understood and limiting its application in safety-sensitive engineering scenarios;3)neglect of causal relationships between variables.Addressing these issues and investigating methods to identify correlations and causal relationships among variables in data-driven transient stability assessment for power systems is of great significance.The specific work carried out in this paper is as follows:1)In terms of characterizing the spatio-temporal correlation between variables in transient stability assessment,two approaches have been proposed.Firstly,a transient stability margin prediction model based on long short-term memory(LSTM)and attention mechanism is proposed,which uses LSTM to extract the temporal features of power flow variables,and adds an attention layer to the traditional deep neural network(DNN)structure to reveal the attention differences of the model on different input variables.Additionally,a transient stability assessment model based on cascading graph and time-convolutional neural network(CNGAT)is proposed,which utilizes the spatial correlation feature representation of input variables generated by the graph attention network’s graph and topology processing ability,and the temporal correlation feature representation of variables generated by the causal dilated convolutional network’s time convolutional feature.The proposed transient stability assessment model effectively extracts and characterizes the spatial-temporal correlation between variables,which further improves the prediction performance of the model and enhances its robustness to topology changes.2)Regarding the characterization of the correlation strength among variables in transient stability evaluation,the gradient-weighted class activation mapping(Grad-CAM)and maximal activation methods are proposed to construct heatmaps and maximal activation maps of the input features for the transient stability evaluation deep learning model.Graph visualization methods are used to characterize the correlation strength between input and output variables,revealing which input features the neural network focuses more on and which inputs can maximize the transient stability classification of the neural network.Additionally,a transient stability evaluation attribution method based on the SHAP framework is proposed to quantitatively evaluate the contribution of different input features to the transient stability evaluation results,thereby guiding the adjustment of power flow in the power grid and improving the transient stability of the power system.3)A data-driven method for identifying causality between transient stability variables in power systems is proposed.Firstly,based on the PC-IGCI algorithm and power system operation dataset,the causal structure among transient stability evaluation variables is discovered,avoiding the Markov equivalence class problem existing in traditional constraint-based causal structure discovery methods.Secondly,a causal effect inference method based on the average causal effect(ACE)and relative average causal effect(RACE)indicators is proposed to quantitatively evaluate the strength of causal relationships between variables.Finally,the causality support rate(CSR)and degree of directional asymmetry(DDA)metrics are proposed to evaluate the reliability of causality relationships using the constructed subset of power grid operation data,thereby revealing the impacts of temporal changes in power grid operating conditions on the causality relationships between transient stability variables in power systems.This research focuses on data-driven transient stability assessment for power systems and proposes methods for identifying and characterizing spatial-temporal correlation features among input variables,attributing relationships between input and output variables,and recognizing causal relationships among variables.These methods address the shortcomings of existing transient stability assessment models in extracting spatial-temporal features,surpassing the current limitations of data-driven approaches limited to correlation mining.The research is of significant importance in understanding the mechanisms of transient stability in power systems and ensuring the secure operation of power grids.