LSTM DRIVEN WATER QUALITY FORECASTING IN UPPER INDUS BASIN

Abstract

This study evaluates the predictive capability of the Long Short-Term Memory (LSTM) model in forecasting water quality parameters of the Swat River at the Chakdara monitoring station using historical data from 1963 to 2020. The dataset was divided into training (70%) and testing (30%) subsets, and the model was developed and optimized using TensorFlow and Keras in Python. Feature engineering and hyperparameter tuning were applied to improve predictive performance for five key water quality indicators: Sodium (Na), Magnesium (Mg), Bicarbonate (HCO₃), Sulfate (SO₄), and Sodium Adsorption Ratio (SAR). The LSTM model achieved excellent results, with an R² of 0.95 (training) and 0.86 (testing), mean squared error (MSE) of 0.003 (training) and 0.008 (testing), Nash-Sutcliffe efficiency (NSE) of 0.93 (training) and 0.86 (testing), and root mean squared error (RMSE) values of 0.05 and 0.09, respectively. These results highlight the LSTM model’s effectiveness in capturing long-term dependencies in hydrological time series, making it highly suitable for water quality prediction. The study underscores the potential of LSTM for integration into environmental monitoring systems to improve the reliability of water quality forecasting and support sustainable water resource management.