Understanding the water flow behavior on an anisotropic wetting surface is of practical significance in nanofluidic devices for their performance improvement. However, current methods of experiments and simulations face challenges in measuring water transportation in real time and visually displaying it. Here, molecular dynamics simulation was integrated with our developed multi-attribute point cloud dataset and a customized network of deep learning to achieve mapping from an anisotropic wetting surface to the static and dynamic behaviors of water molecules and realize the high-performance prediction of water transport behavior. More importantly, for the chaotic phenomenon of water molecule flow caused by thermal fluctuation and limited sampling, we proposed a nanoparticle tracking optimization strategy to improve the prediction performance of the velocity field. The prediction results proved that the deep learning framework proposed in this work had superior performance in terms of accuracy, computational cost and visualization, and had the potential for generality to model the transport behavior of different molecules. Our framework can be expected to motivate the development of real-time water flow prediction at an interface and contribute to the optimization and design of surface structures in nanofluidic devices.