Improving multi-month streamflow forecast in CONUS using Long-Short Term Memory (LSTM) networks

Abstract:

Reliable seasonal streamflow forecasts provide important information on decision-making for a wide range of water resources applications and can be considered a risk indicator in a risk assessment framework. These forecasts are generally challenging due to large temporal and spatial variability of climate. The standard approach of generating seasonal forecasts is to force hydrological models with climate model forecasts (dynamical approach) or to develop a statistical relationship between large-scale climate indices and streamflow (statistical approach). The commonly used dynamical approach suffers from the unreliability of climate models forecasts, which are biased and have considerable uncertainty. In this study, we utilized a hybrid statistical-dynamical approach using a deep learning network, Long Short-Term Memory (LSTM), and climate forecasts of General Circulation Models (GCMs) from the North American Multi-Model Ensemble (NMME). The LSTM network, trained directly with GCM outputs, accounts for the bias within the network and therefore does not require the preprocessing of inputs for bias. Further, we explored the enhancement in forecast capability of our approach by integrating GCM forcings with large-scale climate indices such as Pacific-North America (PNA), North Atlantic Oscillation (NAO), El Niño-Southern Oscillation (ENSO), Arctic Oscillation (AO) and Pacific Decadal Oscillation (PDO). We tested our approach to model the seasonal flow quantiles (from low to high flow) in over 1500 gage stations from the GAGES-II database across CONUS.