Improved runoff forecasting performance through error predictions using a deep-learning approach

Abstract

Accurate runoff prediction is critical for various fields of hydrology, agriculture, and environmental studies. Numerous hydrologic models have been developed and demonstrate good performances in runoff simulation. However, errors are inherent in forecasted runoff predictions, which can cause uncertainty in real-time flood warning systems. In order to improve the predictive performance of hydrologic modeling, this study used a deep learning approach as a post-processor to correct for errors associated with hydrologic data. The proposed model uses the long short-term memory model with sequence-to-sequence structure as a post-processor to improve runoff forecasting. Specifically, the deep learning approach was used to estimate errors in forecasted hourly runoff provided from National Water Model in Russian River basin, California, United States. Error prediction in hourly runoff with lead times between 1 and 18 h were developed using observed precipitation and errors from upstream stream gages to improve the predictive performance of National Water Model. The predictive performance of the model was evaluated using numerous statistical metrics, and results show that the long short-term memory model with sequence-to-sequence post-processor improved runoff predictions compared to standalone results from the National Water Model. Statistical values of percent bias decreased from a range of −60%–80% to −15%–10% when the post-processor model was used, and similarly root mean square errors of runoff prediction decreased from 120 cms to 20 cms. Thus, this study demonstrates the power of deep learning model to improve hydrologic modeling results, especially those with short forecasting lead times.