Development of specialized LSTM runoff forecasting models based on GIS and the SCS–CN method in a complex mountain basin

Abstract

Hydrological forecasting in complex mountain basins is a challenging task. The performance of machine learning forecasting models can be improved by exploiting available spatial rainfall datasets and by incorporating process-based hydrological knowledge, both using feature engineering (FE) strategies. In this study, we assessed the improvement in developing short-term runoff forecasting models using the long short-term memory (LSTM) networks. The selected FE strategies were based on GIS and the Soil Conservation Service Curve Number (SCS-CN) method. To demonstrate the usefulness of the selected FE strategies, we developed referential and specialized (with FE strategies) models for a 3390-km2 basin using the GSMaP-NRT satellite precipitation product (SPP). We developed forecasting models for lead times of 1, 6, and 11 h to account for near-real-time forecasting, flash floods, and concentration time of the basin, respectively. Our results show that the proposed FE strategies improved the efficiencies of LSTM referential models for all lead times, with Nash-Sutcliffe efficiency values of 0.93 (1 h), 0.77 (6 h), and 0.67 (11 h). The utility of the developed models exploiting non-validated satellite precipitation is demonstrated because the results are comparable to those of other studies using ground precipitation information. The proposed methodology and insights from this study provide hydrologists with new tools for developing advanced data-driven runoff models that integrate available geographic information into other precipitation-ungauged hydrological systems.