Comparing Process-Based and Machine Learning Models for Streamflow Prediction in the Kaligandaki River Basin, Nepal

Aayush Roka , Bishesh Khanal

Reliable daily streamflow prediction is critical for hydropower operations, flood risk management, and irrigation planning in monsoon-dominated Himalayan river basins. While both process-based and machine learning (ML) approaches have been used for such tasks, systematic comparisons that decompose the sources of performance differences remain scarce. This study evaluates seven configurations: a process-based SWAT model, and three XGBoost (XGB) and three Random Forest (RF) models representing pure rainfall-runoff, observed-lag, and recursive-simulation scenarios for streamflow prediction in the Kaligandaki river basin, Nepal. An NSE gap decomposition framework is applied to quantify two distinct components of discharge lag information value: the watershed memory benefit and the recursive error propagation cost. During the five-year independent test period, SWAT achieved NSE = 0.851, while pure rainfall runoff XGB-A and RF-A models achieved NSE = 0.840. Observed-lag upper-bound models reached NSE values above 0.94. RF recursive simulation (RF-C) achieved NSE = 0.861, exceeding SWAT, whereas XGB recursive simulation showed no improvement (XGB-C = 0.838), revealing a strong algorithm-dependent sensitivity to recursive error propagation. Flow duration curve analysis reveals that SWAT underestimates low flows (>60% exceedance probability) despite near-zero total PBIAS (−0.28%), reflecting compensating biases between high- and low-flow regimes. SHAP analysis identifies the Antecedent Precipitation Index (API) as the dominant precipitation predictor, with 30-day cumulative rainfall as the second-ranked feature, confirming multi-week soil-moisture memory as the primary catchment-scale control on discharge in this large, monsoon-dominated basin. These results establish that well-engineered pure rainfall-runoff models match SWAT in aggregate NSE while substantially outperforming it in distributional flow reproduction.