Abstract: The subsequent commutation failure of HVDC transmission is difficult to be predicted due to the complex influence factors and high randomness. Extreme learning machine（ELM）is used to predict commutation failure,and a data-driven method for predicting subsequent commutation failure is proposed.The data of inverter-side commutation bus voltage,DC current and trigger delay angle after the first phase change failure are gathered,and 11 fault features such as maximum,average and minimum values of DC current are calculated as the feature samples of ELM classifier. The influence of implicit layer activation function and the number of hidden layer nodes on the accuracy of the model is compared,and a pre-judgment model of subsequent commutation failure is constructed. PSCAD/EMTDC is used to build a high-voltage DC transmission model,train and test are made to the model. The test results verify the effectiveness of the proposed model.