Abstract: Aimed at the issue of hydrological modeling for the areas without or lacking hydrological data, this paper develops an intelligent hydrological model based on the transfer learning and long short-term memory neural network (LSTM) technology, through a case study of the Jialing River, Wujiang River, and Minjiang River basins . First, based on the measured hydrological and meteorological data, this model adopts the K-nearest neighbor algorithm to simulate meteorological data, and uses the soil and water assessment tool (SWAT) to generate river runoff process, creating measured and simulated sample sets. Then, we design different network transfer fine-tuning strategies and network learning scenarios, examine the possibility of network transfer, and compare the performance and effectiveness of the transferred networks. Results show the effect of transfer learning is better when the cell layer of the network is fixed and the other layers are fine-tuned. The transferred network is more stable and more accurate in the case of the network transfer performed in the same watershed or across watersheds. In cross-watershed transfer learning, higher similarity between the source and target watersheds leads to less difficult work and a higher accuracy in network transfer. Thus, this model is a new idea useful for hydrological model development for the areas without or lacking hydrological data.

Key words: transfer learning, long short-term memory neural network, hydrological model, runoff simulation, fine-tuning strategies