Authors: Mostafa Saberian, and Vidya Samadi – Clemson University, USA; Ioana Popescu – TU Delft, the Netherlands
Title: Uncertainty Quantification of Deep Neural Networks for Flood Prediction
Presentation Type: Poster Presentation
Abstract: Effectively characterizing uncertainty and error in flood prediction is essential for informed decision-making. This study combines advanced deep neural network architectures, i.e., Neural Basis Expansion Analysis for Interpretable Time Series Forecasting (N-BEATS), Neural Hierarchical Interpolation for Time Series Forecasting (N-HiTS), and Long Short-Term Memory (LSTM), with multiple uncertainty quantification frameworks to evaluate flood prediction across several watersheds in the southeastern United States. Bayesian inference, Monte Carlo–based methods, and quantile regression are applied to estimate predictive uncertainty. The comparative analysis examines how different uncertainty approaches perform across a range of flood magnitudes, highlighting their respective advantages and limitations at multiple scales. The results show that prediction skill is high across horizons, but uncertainty quantification differed among methods and case studies. Uncertainty quantification parameters mainly control interval uncertainty width, with higher dropout rates or prior standard deviation widening prediction bands but providing limited coverage improvement. Quantile-based formulations generally produce the most reliable intervals. These insights enhance our understanding of combining deep neural networks and uncertainty quantification to reduce prediction errors and improve simulations.