Evaluation and Interpretation of Runoff Forecasting Models Based on Hybrid Deep Neural Networks

Deep neural networks has been widely used in runoff forecasting and has achieved better performance than of conceptual hydrological models. However, most existing studies only use a single type of neural network model to build runoff forecasting models, which fails to fully explain the role of different types of neural networks in runoff forecasting. In this study, the convolutional neural networks (CNN), long short-term memory (LSTM) networks, and convolutional LSTM (ConvLSTM) were used to design a hybrid deep neural network model (HydroDL) for runoff prediction by referring to the structure of the conceptual hydrological model. The proposed model was used to predict the daily runoff at the Xinlong and Daofu hydrological stations in the upper reaches of the Yalong River, and several other runoff prediction models based on neural networks and conceptual hydrological models were developed for comparative study. Daily scale meteorological, hydrological and topographic data from January 2011 to December 2020 were used to train and validate the above models. The results show that: (1) the proposed HydroDL model has higher prediction accuracy and more stable prediction performance than runoff prediction models based on a single type of neural network. (2) effects of different parts in the HydroDL model are distinct, among which the terrain feature extraction and runoff conversion has the most significant and least significant effect, respectively, on improving the forecast results.