Improving Streamflow Simulations using Long Short-Term Memory Networks

With extreme flooding events causing catastrophic damage in India every year, accurate hydrologic predictions are of both scientific and societal significance. Machine Learning models have shown promise in improving streamflow simulations but often require longer training data sets to accurately capture the dynamics of complex hydrologic processes. This study adopts a post-processing methodology entailing feeding the output of a process-based model into a data-driven model. We develop a long-term national-scale streamflow reanalysis product over India and post-process it using LSTMs in order to improve basin-scale simulations. The goal is to perform systematic corrections to basin-scale streamflow simulations using a theory-guided machine learning approach.