Abstract: In order to solve the problems of imbalanced data and poor model interpretability in the state evaluation of UHVDC converter valve, a new evaluation method based on light gradient boosting machine (LightGBM) and SHAP was proposed. Firstly, balanced samples were generated through hierarchical clustering, adaptively determining sub-cluster size and weighted oversampling to solve the problem of sample imbalance. Secondly, a state evaluation model based on the LightGBM tree structure classifier was built to achieve rapid and accurate evaluation of samples. Thirdly, an influencing factors analysis framework for UHVDC converter valve state evaluation based on SHAP attribution theory was presented, which could explain the importance of each state quantity and its impact on the result from the overall and individual perspectives. Through calculation examples, this paper verified the effectiveness of the proposed oversampling method and state evaluation model. The analysis of key influencing factors provided a basis and support for the state evaluation results of the converter valve.