<p>Several deep learning (DL) based river streamflow forecasting studies have been conducted in recent years, with many producing reasonably accurate predictions. Providing uncertainty quantification (UQ) for any forecasting is important, especially in high-risk scenarios like flood forecasting. The novelty of this work lies in its focus on the early phases of flood forecasting and the assessment of uncertainties in both the timing and accuracy of predictions. This study evaluates two uncertainty quantification (UQ) techniques-Monte Carlo dropout and Ensemble methods-with deep learning (DL) models for short-term flood forecasting in the Prüm and Kyll River basins, western Germany. Three DL architectures (LSTM, GRU, and 1D-CNN) are assessed. While both the models and UQ methods perform reasonably well under normal conditions (e.g., average <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11069_2025_7527_Article_IEq1.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="93" /> </InlineMediaObject> <EquationSource Format="TEX">$NSE> 0.85$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>N</mi> <mi>S</mi> <mi>E</mi> <mo>></mo> <mn>0.85</mn> </mrow> </math></EquationSource> </InlineEquation>, MAE <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11069_2025_7527_Article_IEq2.gif" Format="GIF" Height="20" Rendition="HTML" Resolution="72" Type="Linedraw" Width="78" /> </InlineMediaObject> <EquationSource Format="TEX">$< 3.5\, \textrm{m}^3/\textrm{s}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo><</mo> <mn>3.5</mn> <mspace width="0.166667em" /> <msup> <mtext>m</mtext> <mn>3</mn> </msup> <mo stretchy="false">/</mo> <mtext>s</mtext> </mrow> </math></EquationSource> </InlineEquation>), they face significant challenges during the early flood phases. Predictions during these phases are inaccurate, particularly for LSTM and GRU models (e.g., <InlineEquation ID="IEq3"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11069_2025_7527_Article_IEq3.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="93" /> </InlineMediaObject> <EquationSource Format="TEX">$NSE < 0.40$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>N</mi> <mi>S</mi> <mi>E</mi> <mo><</mo> <mn>0.40</mn> </mrow> </math></EquationSource> </InlineEquation>), and the UQ methods exacerbate the issue by producing overly narrow uncertainty intervals, thereby underestimating their uncertainty. The Ensemble method performs better overall, offering narrower and more confident intervals, but all approaches struggle to capture rapid changes in early flood stages, signaling a need for caution in their application.</p>

How confident and reliable are deep learning models for streamflow prediction under flood conditions

期刊论文
发表日期：

2025年7月30日

查看全文

Mohammed Albared,
Hans-Peter Beise,
Manfred Stüber

摘要

Several deep learning (DL) based river streamflow forecasting studies have been conducted in recent years, with many producing reasonably accurate predictions. Providing uncertainty quantification (UQ) for any forecasting is important, especially in high-risk scenarios like flood forecasting. The novelty of this work lies in its focus on the early phases of flood forecasting and the assessment of uncertainties in both the timing and accuracy of predictions. This study evaluates two uncertainty quantification (UQ) techniques-Monte Carlo dropout and Ensemble methods-with deep learning (DL) models for short-term flood forecasting in the Prüm and Kyll River basins, western Germany. Three DL architectures (LSTM, GRU, and 1D-CNN) are assessed. While both the models and UQ methods perform reasonably well under normal conditions (e.g., average $NSE> 0.85$ $N S E > 0.85$ , MAE $< 3.5\, \textrm{m}^3/\textrm{s}$ $< 3.5 m^{3} / s$ ), they face significant challenges during the early flood phases. Predictions during these phases are inaccurate, particularly for LSTM and GRU models (e.g., $NSE < 0.40$ $N S E < 0.40$ ), and the UQ methods exacerbate the issue by producing overly narrow uncertainty intervals, thereby underestimating their uncertainty. The Ensemble method performs better overall, offering narrower and more confident intervals, but all approaches struggle to capture rapid changes in early flood stages, signaling a need for caution in their application.