Scale effects of the monthly streamflow prediction using a state-of-the-art deep learning model

Authors: XU, WENXIN - Wuhan University; CHEN, JIE - Wuhan University; Zhang, Xunchang

Submitted to: Water Resources Management
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 6/4/2022
Publication Date: 6/21/2022
Citation: Xu, W., Chen, J., Zhang, X.J. 2022. Scale effects of the monthly streamflow prediction using a state-of-the-art deep learning model. Water Resources Management. https://doi.org/10.1007/s11269-022-03216-y.
DOI: https://doi.org/10.1007/s11269-022-03216-y

Interpretive Summary: The accurate prediction of monthly streamflow in a watershed or drainage basin is important in sustainable water resources planning and management in the face of the increased variability in precipitation. With the rapid development in Artificial Intelligence (AI) with the advances in computer sciences, there is a growing interest in applying AI (specifically machine learning and deep learning) to surface hydrology prediction. However, two questions related to predictive ability need to be asked when using a deep learning computer model for monthly streamflow prediction: (1) is the predictive performance dependent on the watershed area? and (2) how sensitive is the predictive performance to training data length? To answer these questions, a hybrid deep learning computer model (combining Convolutional Neural Network and Gated Recurrent Unit) was applied for a large number of watersheds with varying hydroclimatic characteristics around the globe. The spatial and temporal scale effects on predictive performance was then investigated for the monthly streamflow prediction. The results show that the deep learning computer model is more suitable for monthly streamflow predictions over watersheds with large surface drainage areas. The surface area of 3,000 square km can be considered as a threshold for the predictive performance. In addition, the predictive performance tends to get better with longer training periods or greater training data. The 25- to 35-year length is the minimum training period to obtain stable predictive performance for most watersheds. The findings should be useful to hydrologists and water resources management engineers for developing AI-based prediction and decision support tools.

Technical Abstract: The accurate prediction of monthly streamflow is important in sustainable water resources planning and management. There is a growing interest in the development of deep learning models for monthly streamflow prediction with the advances in computer sciences. However, two questions related to predictive performance need to be answered when using the deep learning model for monthly streamflow prediction: (1) is the predictive performance dependent on the watershed area? and (2) how sensitive is the predictive performance to training data length? To this end, a hybrid deep learning prediction model combining Convolutional Neural Network and Gated Recurrent Unit (i.e., CNN-GRU) was first proposed and applied for a large number of watersheds with varying hydroclimatic characteristics around the globe. The spatial and temporal scale effects on predictive performance was then investigated for the monthly streamflow prediction. The Nash-Sutcliffe efficiency coefficient (NSE) and mean relative error (MRE) are used as criteria to evaluate the predictive performance. The results show that the deep learning model is more suitable for monthly streamflow predictions over watersheds with large surface areas. The surface area of 3,000 square km can be considered as a threshold for the predictive performance. The median NSE increases from 0.31 to 0.40, while the median MRE decreases from 53.2% to 46.2% for watersheds with areas being larger than 3,000 square km compared with those with areas being smaller than 3,000 square km. In addition, the predictive performance tends to get better with the extension of the training period. When the length of the training period increases stepwise from 10 years to 50 years, there is a large increase in NSE (from 0.28 to 0.40) and a moderate decrease in MRE (from 50.3% to 46.2%) for watersheds with areas being larger than 3,000 square km. Similar changes can also be found for watersheds with areas being smaller than 3,000 square km. The 25- to 35-year is the minimum training period length to obtain stable predictive performance for most watersheds.