Application of Deep Learning Method for Daily Streamflow Time-Series Prediction: A Case Study of the Kowmung River at Cedar Ford, Australia

Abstract

Sustainable management of water supplies faces a comprehensive challenge due to global climate change. Improving forecasts of streamflow based on erratic precipitation is a significant activity nowadays. In recent years, the techniques of data-driven have been widely used in the hydrological parameter’s prediction especially streamflow. In the current research, a deep learning model namely Long Short-Term Memory (LSTM), and two conventional machine learning models namely, Random Forest (RF), and Tree Boost (TB) were used to predict the streamflow of the Kowmung river at Cedar Ford in Australia. Different scenarios proposed to determine the optimal combination of input predictor variables, and the input predictor variables were selected based on the auto-correlation function (ACF). Model output was evaluated using indices of the root mean square error (RMSE), and the Nash and Sutcliffe coefficient (NSE). The findings showed that the LSTM model outperformed RF and TB in predicting the streamflow with RMSE and NSE equal to 102.411, and 0.911 respectively. for the LSTM model. The proposed model could adopt by hydrologists to solve the problems associated with forecasting daily streamflow with high precision. This study may not be generalized because of the geographical condition and the nature of the data for each location.

A Systematic Review of Deep Learning Applications in Streamflow Data Augmentation and Forecasting

“…The rest of the studies from 2021 for daily streamflow modeling were mostly LSTM focused. In mostly basin-specific studies, Latif and Ahmed (2021) 2021) employed SSA and Seasonal-Trend decomposition using LOESS (STL) for the Nallihan River, Turkey. Their approach coupled SSA and STL with ANNs, LSTMs, and CNNs to compare them to vanilla versions of those networks.…”

“…Streamflow forecasting is a crucial component of many issues in hydrology and water management, such as watershed management, agricultural planning (Yildirim and Demir, 2022), flood prediction (Krajewski et al, 2021), and many other mitigation needs (Ahmed et al, 2021;Yaseen et al, 2018;Yildirim and Demir, 2021). However, accurate and reliable forecasting is challenging due to the complexity of hydrological systems, including nonlinearity and dynamic behavior as well as randomness in the datasets (Honorato et al, 2018;Yaseen et al, 2017).…”

Application of data-driven models to predict the dimensions of flow separation zone

“…These methods are well known as the top adaptive ones suitable for finding complex and nonlinear indefinite patterns in large dimensional data. As scientists’ skill with these AI-based systems deepens, they are becoming more dependable, and now they are frequently utilized as robust approaches in different fields of water sciences to predict complex hydraulic and hydrological variables such as sugarcane growth based on climatological parameters (Taherei Ghazvinei et al 2018 ), daily dew point temperature (Qasem et al 2019 ), forecasting nitrate concentration as a water quality parameter (Latif et al 2020 ), inflow forecasting (Latif et al 2021a ), phosphate forecasting in reservoir water system (Latif et al 2021b ), daily streamflow time-series prediction (Latif and Ahmed 2021 ; Tofiq et al 2022 ), surface water quality status and prediction during movement control operation order under COVID-19 pandemic (Najah et al 2021 ), groundwater level fluctuations (Ghasemlounia et al 2021 ; Gharehbaghi et al 2022 ), discharge coefficient of a new type of sharp-crested V-notch weirs (Gharehbaghi and Ghasemlounia 2022 ), and dissolved oxygen prediction (Ziyad Sami et al 2022 ).…”