Scholarlink

<p>Reliable discharge data are essential for hydrological modeling, flood forecasting, and water resources management, yet many river basins in India experience prolonged data gaps due to gauge failures and sedimentation effects. This study develops a physics-guided hybrid framework to reconstruct missing discharge records at the Hajipur station of the Gandak River, a hydraulically complex reach influenced by backwater conditions near its confluence with the Ganga. A cross-section–based hydraulic model grounded in the Manning–Strickler formulation and calibrated using detailed channel geometry with piecewise conveyance coefficients for low- and high-flow regimes was first established to provide a physically consistent discharge baseline. The physics-based model showed moderate efficiency (R<sup>2</sup> = 0.57; NSE = 0.56), indicating systematic magnitude deviations under dynamic flow conditions. Residual deviations were subsequently modeled using Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Gradient Boosting (GBR) using local stage and upstream discharge information. Chronological validation (75–25 split) revealed distinct differences in generalization performance. While RF and GBR achieved high calibration accuracy, their validation efficiency declined substantially (NSE ≤ 0.38), indicating overfitting. In contrast, XGBoost maintained strong validation performance (R<sup>2</sup> = 0.92; NSE = 0.90) with markedly reduced error magnitude compared to the physics baseline. The hybrid XGBoost framework effectively corrected systematic hydraulic bias while preserving physically realistic discharge responses. Overall, the proposed physics–machine learning integration provides a robust and scalable strategy for discharge reconstruction in data-limited and hydrodynamically complex river systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A physics-guided machine learning framework for accurate reconstruction of missing discharge records: case study of the Gandak River at Hajipur, India

  • 期刊论文

  • 发表日期：

    2026年5月22日

查看全文
  • Rahul Prakash,
  • Joseph Tripura

摘要

Reliable discharge data are essential for hydrological modeling, flood forecasting, and water resources management, yet many river basins in India experience prolonged data gaps due to gauge failures and sedimentation effects. This study develops a physics-guided hybrid framework to reconstruct missing discharge records at the Hajipur station of the Gandak River, a hydraulically complex reach influenced by backwater conditions near its confluence with the Ganga. A cross-section–based hydraulic model grounded in the Manning–Strickler formulation and calibrated using detailed channel geometry with piecewise conveyance coefficients for low- and high-flow regimes was first established to provide a physically consistent discharge baseline. The physics-based model showed moderate efficiency (R2 = 0.57; NSE = 0.56), indicating systematic magnitude deviations under dynamic flow conditions. Residual deviations were subsequently modeled using Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Gradient Boosting (GBR) using local stage and upstream discharge information. Chronological validation (75–25 split) revealed distinct differences in generalization performance. While RF and GBR achieved high calibration accuracy, their validation efficiency declined substantially (NSE ≤ 0.38), indicating overfitting. In contrast, XGBoost maintained strong validation performance (R2 = 0.92; NSE = 0.90) with markedly reduced error magnitude compared to the physics baseline. The hybrid XGBoost framework effectively corrected systematic hydraulic bias while preserving physically realistic discharge responses. Overall, the proposed physics–machine learning integration provides a robust and scalable strategy for discharge reconstruction in data-limited and hydrodynamically complex river systems.