Executive Summary : | Hydrological processes leading to flow in river channels are complex, making it difficult to find a fundamental law to explain river flow dynamics at basin-scale. Hydrological models typically have free-parameters that need to be calibrated using observed discharge time series data. However, attempts to relate observable catchment characteristics with model parameters have been unsuccessful due to structural uncertainties and the inability to measure many characteristics of natural river basins. To predict discharge in discharge-data scarce river basins, hydrologists have devised regionalization methods, which rely on transforming relevant information from gauged basins to hydrological similar ungauged basins. However, these methods are not reliable if the density of discharge-gauging stations in the region is low. This proposed research hypothesizes that assuming all river basins are hydrologically similar is not expected to reduce model performance. A dynamic hydrological model based on the Budyko curve concept has been developed, but its performance is not up to the mark. Machine learning model-based studies, particularly those based on LSTM, have shown that superior prediction can be achieved by adopting appropriate machine learning structures.
The aim of this proposal is to combine the best of DB and LSTM approaches, using a global dataset with nearly 4000 river-discharge data to ensure the hydrological model's reliability in data-scarce regions. |