Abstract: Aiming at the issue about online prediction of the complicated nonlinear time series, an online prediction method based on process neural network (PNN) was proposed. The prediction model of double parallel feedforward discrete input process neural network (DPFDPNN), which was trained by off-line data, was firstly established to make prediction of the complicated nonlinear time series. In order to improve the accuracy and efficiency of the DPFDPNN for the time series prediction, the weight connecting, the hidden layer and output layer were then directly updated by the recursion extreme learning algorithm (REL) based on recursive algorithm with the real data stream. Finally, the DPFDPNN with the updated weight was adopted to predict the time series. The corresponding learning algorithm for DPFDPNN and the updating mechanisms were obtained in this paper. The prediction method mentioned above was verified by a test case with chaotic time series, and an example of method application in condition pre-diction of liquid propellant rocket engine was given. The results show that the proposed method outperformed the DPFDPNN without weight updated at accuracy and adaptability. It’s an effective way to solve the failure online prediction problem of the complicated nonlinear time series.