Intercomparison of deep learning models in predicting streamflow patterns: insight from CMIP6

Abstract

Abstract This research was carried out to predict daily streamflow for the Swat River Basin, Pakistan through four deep learning (DL) models: Feed Forward Artificial Neural Networks (FFANN), Seasonal Artificial Neural Networks (SANN), Time Lag Artificial Neural Networks (TLANN) and Long Short-Term Memory (LSTM) under two Shared Socioeconomic Pathways (SSPs) 585 and 245. Taylor Diagram, Random Forest, and Gradient Boosting techniques were used to select the best combination of General Circulation Models (GCMs) for Multi-Model Ensemble (MME) computation. MME was computed via the Random Forest technique for Maximum Temperature (Tmax), Minimum Temperature (Tmin), and precipitation for the aforementioned three techniques. The best MME for Tmax, Tmin, and precipitation was rendered by Compromise Programming. The DL models were trained and tested using observed precipitation and temperature as independent variables and discharge as dependent variables. The results of deep learning models were evaluated using statistical performance indicators such as root mean square error (RMSE), mean square error (MSE), mean absolute error (MAE), and coefficient of determination (R2). The TLANN demonstrated superior performance compared to the other models based on RMSE, MSE, MAE, and R2 during training (65.25 m3/s, 4256.97 m3/s, 46.793 m3/s and 0.7978) and testing (72.06 m3/s, 5192.95 m3/s, 51.363 m3/s and 0.7443) respectively. Subsequently, TLANN was utilized to make predictions based on MME of SSP245 and SSP585 scenarios for future streamflow until the year 2100. These results can be used for planning, management, and policy-making regarding water resources projects in the study area.

Keywords

• Streamflow
• Prediction
• Deep learning
• GCM
• CMIP6