Performance Comparison of an LSTM-based Deep Learning Model versus Conventional Machine Learning Algorithms for Streamflow Forecasting

Streamflow forecasting plays a key role in improvement of water resource allocation, management and planning, flood warning and forecasting, and mitigation of flood damages. There are a considerable number of forecasting models and techniques that have been employed in streamflow forecasting and gained importance in hydrological studies in recent decades. In this study, the main objective was to compare the accuracy of four data-driven techniques of Linear Regression (LR), Multilayer Perceptron (MLP), Support Vector Machine (SVM), and Long Short-Term Memory (LSTM) network in daily streamflow forecasting. For this purpose, three scenarios were defined based on historical precipitation and streamflow series for 26 years of the Kentucky River basin located in eastern Kentucky, US. Statistical criteria including the coefficient of correlation ( $$R$$ R ), Nash-Sutcliff coefficient of efficiency ( $$E$$ E ), Nash-Sutcliff for High flow ( $${E}_{H}$$ E H ), Nash-Sutcliff for Low flow ( $${E}_{L}$$ E L ), normalized root mean square error ( $$NRMSE$$ NRMSE ), relative error in estimating maximum flow ( $$REmax$$ REmax ), threshold statistics ( $$TS$$ TS ), and average absolute relative error ( $$AARE$$ AARE ) were employed to compare the performances of these methods. The results show that the LSTM network outperforms the other models in forecasting daily streamflow with the lowest values of $$NRMSE$$ NRMSE and the highest values of $${E}_{H}$$ E H , $${E}_{L}$$ E L , and $$R$$ R under all scenarios. These findings indicated that the LSTM is a robust data-driven technique to characterize the time series behaviors in hydrological modeling applications.