Scholarlink

<p>River discharge rate fluctuations due to climate change and anthropogenic activities are important factors to consider for inferring possible changes in basin hydrological characteristics. River discharge periodicity was comprehensively evaluated, using statistical and a hybrid ensemble machine learning (ML) framework across four in situ gauge stations (Muri, Adityapur, Jamshedpur, and Ghatsila) in the Subarnarekha River Basin (SRB) for the period covering the years 1981–2020. Trend analysis using Mann–Kendall (MK), Spearman’s rho, and innovative trend analysis (ITA) revealed spatial heterogeneity, with a significant annual decline at Muri (− 0.60 m<sup>3</sup>/s/year; <i>D</i> = − 3.84) and a mild positive trend at Ghatsila (+ 0.596 m<sup>3</sup>/s/year; <i>D</i> = 0.58). The seasonal analysis identified strong pre-monsoon and winter trends, notably at Ghatsila (<i>Z</i> = 4.509; slope = + 0.815 m<sup>3</sup>/s). To improve discharge forecasting, multiple ML algorithms, RF (random forest), SVR (support vector regressor), GB (gradient boosting), ADB (adaptive boosting), ET (extra trees), and KNN (K-nearest neighbors), were benchmarked against a proposed MultiBase Gradient Regressor (MBGR). MBGR, a stacking ensemble of ET, ADB, RF, and SVR with GB as the meta-learner, demonstrated the most consistent overall performance across stations and evaluation metrics root mean square error&#xa0;(RMSE)—14.33; mean absolute error (MAE)—6.1. Forecasts for 2021–2030 showed mean annual discharge estimates of 10.50 m<sup>3</sup>/s (Muri), 9.49 m<sup>3</sup>/s (Adityapur), 97.93 m<sup>3</sup>/s (Jamshedpur), and 47.02 m<sup>3</sup>/s (Ghatsila), reflecting historical patterns. The feature importance analysis highlighted rainfall and temperature as key drivers, with regional variability: Precipitation contributed 41.2% to Adityapur, while minimum temperature dominated Ghatsila (37.5%) and Muri (39.8%). These findings highlight the importance of developing location-specific models and demonstrate the effectiveness of ensemble learning, integrated with explainable AI in hydrological forecasting. This should provide critical insights toward imminent climate-resilient water resource management.</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

River discharge forecasting using a stacking ensemble framework with feature attribution

  • 期刊论文

  • 发表日期：

    2026年6月16日

查看全文
  • Prashant Parasar,
  • Akhouri Pramod Krishna

摘要

River discharge rate fluctuations due to climate change and anthropogenic activities are important factors to consider for inferring possible changes in basin hydrological characteristics. River discharge periodicity was comprehensively evaluated, using statistical and a hybrid ensemble machine learning (ML) framework across four in situ gauge stations (Muri, Adityapur, Jamshedpur, and Ghatsila) in the Subarnarekha River Basin (SRB) for the period covering the years 1981–2020. Trend analysis using Mann–Kendall (MK), Spearman’s rho, and innovative trend analysis (ITA) revealed spatial heterogeneity, with a significant annual decline at Muri (− 0.60 m3/s/year; D = − 3.84) and a mild positive trend at Ghatsila (+ 0.596 m3/s/year; D = 0.58). The seasonal analysis identified strong pre-monsoon and winter trends, notably at Ghatsila (Z = 4.509; slope = + 0.815 m3/s). To improve discharge forecasting, multiple ML algorithms, RF (random forest), SVR (support vector regressor), GB (gradient boosting), ADB (adaptive boosting), ET (extra trees), and KNN (K-nearest neighbors), were benchmarked against a proposed MultiBase Gradient Regressor (MBGR). MBGR, a stacking ensemble of ET, ADB, RF, and SVR with GB as the meta-learner, demonstrated the most consistent overall performance across stations and evaluation metrics root mean square error (RMSE)—14.33; mean absolute error (MAE)—6.1. Forecasts for 2021–2030 showed mean annual discharge estimates of 10.50 m3/s (Muri), 9.49 m3/s (Adityapur), 97.93 m3/s (Jamshedpur), and 47.02 m3/s (Ghatsila), reflecting historical patterns. The feature importance analysis highlighted rainfall and temperature as key drivers, with regional variability: Precipitation contributed 41.2% to Adityapur, while minimum temperature dominated Ghatsila (37.5%) and Muri (39.8%). These findings highlight the importance of developing location-specific models and demonstrate the effectiveness of ensemble learning, integrated with explainable AI in hydrological forecasting. This should provide critical insights toward imminent climate-resilient water resource management.