Accurate short-term streamflow forecasting is essential for mitigating urban flood risks and ensuring sustainable water resource management. This study presents a comparative analysis of advanced data-driven techniques for streamflow prediction in the Gandak River basin, India, evaluating their efficacy in handling rapid hydrological fluctuations. Three sophisticated machine learning models – random forest (RF), extreme gradient boosting (XGBoost) and long short-term memory (LSTM) networks – are employed to forecast streamflow across lead times ranging from 1 to 30 days. The findings reveal that all models exhibit strong predictive capabilities for 1-day lead times, achieving high coefficients of determination (R2): 0.976 (LSTM), 0.970 (RF) and 0.964 (XGBoost). However, predictive accuracy diminishes significantly as the forecasting horizon extends. By the 30-day lead time, the LSTM network consistently outperforms the tree-based models (R2 = 0.759 compared to RF’s 0.564 and XGBoost’s 0.569). LSTM’s superior performance for medium- to long-term forecasts is attributed to its advanced architecture, which effectively captures long-term dependencies and non-linear sequential patterns in hydrological time-series data. This study demonstrates the added value of deep learning over traditional ensemble methods for extended forecasting, offering actionable insights into optimising predictive frameworks for urban flood preparedness.
Prakash et al. (Wed,) studied this question.