Evaluating best machine learning model for discharge prediction under dam-regulated river using a novel approach of...

Challenges in regulated basins, such as temporal persistence, nonlinear behaviour, extreme skewness, and regime shifts, undermine data-driven models and lead to systematic underestimation of high flows. This study evaluates 17 model configurations across seven groups of architectural advancements for design contribution. Novel hydrological-domain-motivated designs are developed in the study. A unidirectional LSTM encoder combined with causal self masking accounting for temporal irreversibility, a dual task shared architecture involving discharge regression and peak flow classification to address structural peak underestimation, Regime-Aware Loss function at and above Q90, data augmentation specifically at monsoon-condition and flow region greater than Q90 to generate synthetic training samples while preserving the bimodal discharge distribution, and hydro feature engineering on univariate daily discharge data (1972-2022) to create 25 hydro features. The developed framework reveals training design to be a superior determinant of accuracy, contributing an 18% additional NSE gain over the 9% from raw baselines across four external models. The weighted ensemble achieves a mean NSE of 0.636 and KGE of 0.709, simultaneously reducing 63% PBIAS relative to Base LSTM. Independent protocols as employed, a single split, walk-forward cross-validation and seed stability analysis retain 0.04 NSE throughout experiments. Overall, it is found that prediction accuracy primarily depends on discharge distribution rather than on model architecture. Discharge skewness contributes significantly to NSE prediction variability (50.2 to 55.8%, r = -0.708 to -0.747) across all models and results in a performance gap of 0.390 NSE, which is three times the maximum gain achievable through architectural design (0.138 NSE).