In the middle reaches of the Tarim River, there are many branches and serious overflow of water. To address the issue of insufficient flow data caused by the scarcity and distance of hydrometric stations, the flow prediction models based on BP and RBF artificial neural networks were proposed under the fixed bed and movable bed conditions respectively in the main stream of the Tarim River. The results show that both of the two models have good adaptability under the fixed bed condition. The prediction model of four variables (water depth, water surface width, average flow velocity and hydraulic radius) has the highest accuracy, and the accuracy of the model decreases with the decrease of the number of variables. Under the movable bed condition, the accuracy of BP model is higher than that of RBF model. The BP model trained by three or four variables can better predict the flow rate, but the acquisition of three variables (water depth, water surface width, average flow rate) are convenient. This study could provide new ideas and methods for the river flow prediction in the main stream of the Tarim River, and has practical significance for river management.