Abstract: In order to overcome the deficiency of the traditional time series prediction method in flow prediction of natural gas pipeline networks, a combined prediction model based on Empirical Mode Decomposition (EMD), Attention and Gated Recurrent Unit (GRU) was proposed. Specifically, the model is to substitute the raw flow data of the natural gas pipeline network with its time series component obtained through Empirical Mode Decomposition (EMD), input the intrinsic mode function component obtained into the GRU neural network, calculate the attention probability weight at different times with the attention integrated into the network, and finally learn in the network and predict the time series of flow in the natural gas pipeline network. The verification results in a natural gas pipeline network show that: the EMD-Attention-GRU combined model demonstrates remarkable performance in flow prediction of natural gas pipeline network, capable of capturing the complex non-linear relationships. Besides, the average absolute percentage error of prediction by the combined model outperforms the single GRU and Attention-GRU models by 6.29% and 5.17%, respectively. Thus, it is indicated that the EMD-Attention-GRU combined model could better address the complexities and dynamic features of flow in natural gas pipeline networks than the conventional time-series prediction methods, with values for promotion and application.