Innovative knowledge-based system for streamflow hindcasting: A comparative assessment of Gaussian Process-Integrated Neural Network with LSTM and GRU models

Abstract

Lack of historical data is a major bottleneck for hydrologists to proceed with reliable climate change studies. This work proposes Gaussian Process-Integrated Neural Network (GAUSNET) technique for streamflow hindcasting by considering significant hydrological variables and Global climatic oscillations (GCO) identified by Variance Inflation Factor as system inputs. Dynamic Time Warping based Interpolation is utilized to align monthly GCOs with daily streamflows, followed by feature selection and auto-correlation using Gradient Boosting Machines. On applying for streamflow hindcasting of Greater Pamba, Kerala, India, GAUSNET consistently outperformed Long-Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) across all of the input scenarios with an average Nash-Sutcliffe Efficiency (NSE) of 0.93. GAUSNET based hindcasting can overcome issues of data shortage, fill the data gaps and capture extreme events. Moreover, its ability for uncertainty quantification enhances the reliability and make it as robust tool for hydrological modeling, flood risk assessment, and sustainable water management.