Model Description

This project implements a Hybrid Data-Driven and Process-Based Modeling framework for streamflow forecasting. It integrates the physically-based QSWATPLUS model with advanced machine learning algorithms including Random Forest (RF), XGBoost, LightGBM, and LSTM.

The integration is achieved via a Stacking Ensemble (Meta-learner) and Bayesian Model Averaging (BMA), which effectively bridge the gap between physical process understanding and the pattern-recognition capabilities of ML.

Key Features

• Feature Engineering: Incorporation of lagged flow, rolling precipitation statistics, and seasonal time-based features.
• Uncertainty Analysis: Normality testing of residuals and RAPS analysis to identify persistent changes.
• Climate Change Projections: Forecasting under CMIP6 scenarios (SSP245 and SSP585) for near-term and mid-century periods.

Results Summary

• Superior Accuracy: The Hybrid Model (Stacking Ensemble) achieved the best performance (R² ≈ 0.968, NSE ≈ 0.968) compared to individual models.
• Trend Analysis: Mann-Kendall tests revealed a significant increasing trend in daily and seasonal flow for most models.
• Robustness: ML models significantly improved upon the baseline physically-based model by capturing non-linear hydrological dynamics.

The Stacking Ensemble (Hybrid Model), which integrates the physically-based QSWATPLUS model with machine learning algorithms (RF, XGBoost, LightGBM, and LSTM), represents the pinnacle of this project's forecasting framework.

1. Quantitative Metrics (Validation Period 2001-2008)

• R² (Coefficient of Determination): ≈ 0.968
• NSE (Nash-Sutcliffe Efficiency): ≈ 0.968
• RMSE (Root Mean Square Error): ≈ 14.664 m³/s
• PBIAS (Percent Bias): ≈ 0.000%

2. Comparative Advantage

• Superior Accuracy: The Hybrid Model significantly outperformed the standalone QSWATPLUS (NSE ≈ 0.600) and slightly improved upon individual ML models (e.g., RF NSE ≈ 0.962).
• Reduced Uncertainty: Stacking effectively corrected systematic biases, as evidenced by a PBIAS near zero and residuals that are more centered and less skewed compared to individual components.
• Temporal Consistency: Mann-Kendall tests confirmed that while observed flows showed an increasing trend, the hybrid model residuals maintained 'no trend' (p-value > 0.05), indicating consistent performance across the entire validation timeframe.