Advanced machine learning techniques for daily streamflow forecasting: a case study of the Brahmaputra River

Accurate multi-day streamflow forecasts support flood preparedness, but single holdout scores can conceal lead-time dependence, temporal instability, baseline skill, and uncertainty. This study compares Long Short-Term Memory, Random Forest, Multiple Linear Regression, Multi-Layer Perceptron, Support Vector Regression, and AdaBoost using an audited Bangladesh Water Development Board discharge record for the Bahadurabad Transit station on the Brahmaputra River. The source contained 10,314 entries from 1995 to 2023. After 275 excess same-date records were consolidated by date-wise averaging, 553 calendar days were missing; no temporal interpolation was applied. We constructed 9,272 complete samples using 30 antecedent discharges to predict Days 1–30 directly. Leakage-resistant evaluation used three expanding rolling-origin folds, target-end-date training cut-offs, a 2019–2020 calibration period, and an independent 2021–2023 test period. Moving-block bootstrap confidence intervals, split-conformal prediction intervals, flow-regime errors, and lag sensitivity were evaluated. Random Forest had the lowest mean rolling-origin RMSE, whereas Multi-Layer Perceptron had the lowest pooled test RMSE (6,763.47 m³/s; R²/NSE = 0.721); their uncertainty intervals overlapped. Ranking changed with horizon: Multiple Linear Regression was best at Day 1 and AdaBoost at Day 30. All models underestimated the highest-flow decile. The results support an uncertainty-aware discharge-only benchmark, not universal model superiority or operational flood-warning readiness.