Aiming at the forecasting of time series of flow parameters and flow pattern transformation in offshore oil and gas pipelines, a combined model based on empirical mode decomposition (EMD) and long short-term memory (LSTM) neural networks is established, and Bayesian theory is used to optimize the relevant parameters of LSTM neural network. Compared with BP neural network, random forest algorithm, and LSTM neural network alone, the combined EMD-LSTM prediction model proposed in this paper can better track the evolution trend of riser pressure difference and amplitude, and greatly improve the prediction accuracy. Moreover, it is applicable to both an original flow signal and the time series of its statistical parameters.