Scholarlink

<p>Reliable streamflow prediction under uncertainty remains a central challenge in hydrology, particularly across heterogeneous flow regimes. This study compares a conventional artificial neural network (ANN) with a hybrid HEC–HMS–ANN framework that integrates physically based rainfall–runoff simulation with data-driven learning. Dual uncertainty quantification approaches—Monte Carlo (MC) simulation and Ensemble averaging—were employed to assess calibration (training) and validation (testing) performance. Both models demonstrated strong global predictive skill, with the hybrid framework marginally outperforming ANN in terms of accuracy and stability (training: R<sup>2</sup> = 0.98, RMSE = 51.54–51.72&#xa0;m<sup>3</sup>/s; testing: R<sup>2</sup> = 0.97, RMSE = 50.27–50.47&#xa0;m<sup>3</sup>/s). Ensemble averaging yielded near-perfect global performance (R<sup>2</sup>&#xa0;≈&#xa0;0.99) with reduced RMSE, underscoring its superiority over MC for uncertainty mitigation. Flow-regime-specific analysis revealed the highest reliability for medium flows, while low (baseflow) and high (flood) flows remained more uncertain due to observational noise and data sparsity in extreme events. Uncertainty decomposition indicated that aleatoric variability dominated (&gt; 97%), while epistemic contributions were minor (&lt; 3%), confirming that prediction is constrained primarily by input data quality rather than structural model instability. Despite these strengths, the models are not free from limitations. ANN performance is sensitive to data availability and may risk overfitting, while the hybrid framework introduces additional complexity that slightly increases epistemic uncertainty. Moreover, both approaches exhibited reduced skill in capturing baseflow and extreme floods. Overall, the results highlight the practical value of hybridization for robust, uncertainty-aware forecasting, while recognizing the need for improved input data quality and methodological refinements to enhance reliability across all flow regimes.</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

Hybrid data-driven and process-based modeling for streamflow forecasting: comparative uncertainty analysis in the upper Baro watershed

  • 期刊论文

  • 发表日期：

    2025年10月28日

查看全文
  • Yonata Belina,
  • Asfaw Kebede,
  • Muthoni Masinde

摘要

Reliable streamflow prediction under uncertainty remains a central challenge in hydrology, particularly across heterogeneous flow regimes. This study compares a conventional artificial neural network (ANN) with a hybrid HEC–HMS–ANN framework that integrates physically based rainfall–runoff simulation with data-driven learning. Dual uncertainty quantification approaches—Monte Carlo (MC) simulation and Ensemble averaging—were employed to assess calibration (training) and validation (testing) performance. Both models demonstrated strong global predictive skill, with the hybrid framework marginally outperforming ANN in terms of accuracy and stability (training: R2 = 0.98, RMSE = 51.54–51.72 m3/s; testing: R2 = 0.97, RMSE = 50.27–50.47 m3/s). Ensemble averaging yielded near-perfect global performance (R2 ≈ 0.99) with reduced RMSE, underscoring its superiority over MC for uncertainty mitigation. Flow-regime-specific analysis revealed the highest reliability for medium flows, while low (baseflow) and high (flood) flows remained more uncertain due to observational noise and data sparsity in extreme events. Uncertainty decomposition indicated that aleatoric variability dominated (> 97%), while epistemic contributions were minor (< 3%), confirming that prediction is constrained primarily by input data quality rather than structural model instability. Despite these strengths, the models are not free from limitations. ANN performance is sensitive to data availability and may risk overfitting, while the hybrid framework introduces additional complexity that slightly increases epistemic uncertainty. Moreover, both approaches exhibited reduced skill in capturing baseflow and extreme floods. Overall, the results highlight the practical value of hybridization for robust, uncertainty-aware forecasting, while recognizing the need for improved input data quality and methodological refinements to enhance reliability across all flow regimes.