A framework of integrating heterogeneous data sources for monthly streamflow prediction using a state-of-the-art deep learning model

Abstract

Deep learning has been widely used in hydrological prediction such as monthly streamflow and its performance is usually dependent on the abundance of training data. Even though the interest in using predictors from multiple data sources (e.g., streamflow observations, local meteorological data, and large-scale climate indexes) to train deep learning models for monthly streamflow prediction is growing, these predictors are usually selected from historical periods. Such approaches have limitations that the non-stationary future climate information is not included in the deep learning models. Climate models can provide non-stationary climate information for the future period, which may be useful for monthly streamflow prediction. Therefore, the objectives of this study are to (1) investigate the added value of using predictors derived from global climate models (GCMs) for monthly streamflow prediction based on a state-of-the-art deep learning model, and (2) propose a framework for integrating heterogeneous data sources for monthly streamflow prediction. The framework consists of five integral components: data collection, predictor combination, predictor selection, model construction, and results evaluation. In this study, a hybrid deep learning model combining Convolutional Neural Network and Gated Recurrent Unit was applied for six hydrological stations from mainstream and six stations from tributaries of the Yangtze River. Historical hydroclimate data and GCM hindcasts from 1982 to 2010 are used in monthly streamflow forecasts. Hindcasts are retrospective forecasts for many variables in the past, ideally conducted using the same model used for real-time forecasts. The results show that GCM hindcasts are useful predictors to improve the prediction accuracy for monthly streamflow predictions, especially for the 1- and 3-month lead times. Combining GCM hindcasts with either historical meteorological data or historical streamflow observations and meteorological data as predictors generally provides the best predictive performance. In addition, using large-scale climate indexes as ancillary information is able to improve the predictive performance at a lead time of 6 months. For lead times of 1, 3, and 6 months, the Kling‐Gupta efficiency (KGE) and the mean relative error (MRE)) metrics calculated based on the best-performing predictor combinations are satisfactory for hydrological stations in both mainstream and tributaries, with the median KGE being higher than 0.85 and 0.62, and the median MRE being approximately 20 % and 40 %, respectively, suggesting the monthly streamflow predictions are better for mainstream than for tributaries. Overall results show that (1) the inclusion of predicted information from GCMs can improve the performance for monthly streamflow prediction, and (2) the way of combining various heterogeneous data sources is crucial.