Scholarlink

<p>Reliable rainfall forecasting remains essential for effective water resources management and flood risk mitigation. This study develops two hybrid deep learning frameworks: a Convolutional Neural Network combined with Q-learning (CNN–Q-learning) and a Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) to simulate multi-station precipitation across Iran’s Karkheh Basin. Among the first, Evidential Deep Learning (EDL) is embedded into these hybrid models to analytically disentangle aleatoric (data-related) and epistemic (model-related) uncertainties within a single-pass predictive framework. Monthly precipitation data (1966–2019) from five synoptic stations were quality-controlled, normalized via Min–Max scaling, and used to train the models under a strict chronological split. To assess robustness and sampling uncertainty, bootstrap resampling and Latin Hypercube Sampling (LHS) were employed to generate ensemble scenarios and confidence intervals. The results demonstrate that CNN–Q-learning more effectively captures extreme precipitation and heavy-tailed distributions (Mean RMSE ≈ 1.7&#xa0;mm; NSE ≈ 0.78; CRPS ≈ 0.65; KS ≈ 0.055), while CNN–LSTM yields sharper central predictions with slightly higher overall accuracy (Mean RMSE ≈ 1.1&#xa0;mm; NSE ≈ 0.79; CRPS ≈ 0.91). EDL-derived uncertainty components indicate normalized aleatoric and epistemic variances, confirming the model’s ability to quantify both measurement and model uncertainty. Both models achieved full coverage of observations within 95% confidence bounds; however, CNN–Q-learning showed better calibration in the tails, while CNN–LSTM exhibited narrower predictive intervals. Overall, integrating EDL-based uncertainty quantification with an adaptive Q-learning policy provides a transparent, interpretable, and decision-oriented framework for probabilistic rainfall simulation. The CNN–Q-learning model is more suitable for risk management and extreme-event forecasting, whereas CNN–LSTM offers higher precision for routine operational predictions.</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 Novel Evidential Uncertainty Framework for Hybrid Models in Rainfall Simulation

  • 期刊论文

  • 发表日期：

    2026年1月16日

查看全文
  • Hamid Ebrahimi

摘要

Reliable rainfall forecasting remains essential for effective water resources management and flood risk mitigation. This study develops two hybrid deep learning frameworks: a Convolutional Neural Network combined with Q-learning (CNN–Q-learning) and a Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) to simulate multi-station precipitation across Iran’s Karkheh Basin. Among the first, Evidential Deep Learning (EDL) is embedded into these hybrid models to analytically disentangle aleatoric (data-related) and epistemic (model-related) uncertainties within a single-pass predictive framework. Monthly precipitation data (1966–2019) from five synoptic stations were quality-controlled, normalized via Min–Max scaling, and used to train the models under a strict chronological split. To assess robustness and sampling uncertainty, bootstrap resampling and Latin Hypercube Sampling (LHS) were employed to generate ensemble scenarios and confidence intervals. The results demonstrate that CNN–Q-learning more effectively captures extreme precipitation and heavy-tailed distributions (Mean RMSE ≈ 1.7 mm; NSE ≈ 0.78; CRPS ≈ 0.65; KS ≈ 0.055), while CNN–LSTM yields sharper central predictions with slightly higher overall accuracy (Mean RMSE ≈ 1.1 mm; NSE ≈ 0.79; CRPS ≈ 0.91). EDL-derived uncertainty components indicate normalized aleatoric and epistemic variances, confirming the model’s ability to quantify both measurement and model uncertainty. Both models achieved full coverage of observations within 95% confidence bounds; however, CNN–Q-learning showed better calibration in the tails, while CNN–LSTM exhibited narrower predictive intervals. Overall, integrating EDL-based uncertainty quantification with an adaptive Q-learning policy provides a transparent, interpretable, and decision-oriented framework for probabilistic rainfall simulation. The CNN–Q-learning model is more suitable for risk management and extreme-event forecasting, whereas CNN–LSTM offers higher precision for routine operational predictions.