Comparison of multi-DLM approaches for predicting daily runoff: evidence from the data-driven model in one of China's largest wheat production-bases.

Abstract

Runoff forecasting is extremely important for various activities of water pollution research and agricultural. Data-driven models have been proved an effective approach in predicting daily runoff when combining deep learning methods (DLM). However, predicting accuracy of daily runoff still need improved. Here, we firstly proposed a combined model of Gate Recurrent Unit (GRU) and Residual Network (ResNet) and compared with one shallow learning method (Back Propagation Neural Network, BPNN) and one deep learning method (GRU) with data from 2010 to 2020 in three stations in daily runoff forecasting in the Yiluo River watershed. The results showed that the combined model with precipitation data and runoff data as input has the highest prediction accuracy (NSE = 0.9325, 0.8735, 0.9186, respectively). Input data with precipitation have higher prediction accuracy than that without. The performance of the model was better in the dry season than the wet season. The topographic and geomorphic factors may also the main factors affecting runoff forecast. Those results of this study can provide useful strategies to predict short runoff and manage watershed scale water resources especially in the important agriculture region.