Logo image
Back
Uncertainty Quantification for Streamflow Forecasting via Spatio-Temportal Conformal Prediction
Thesis

Uncertainty Quantification for Streamflow Forecasting via Spatio-Temportal Conformal Prediction

Mohammed Amine Gharsallaoui
Master of Science (MS), Washington State University
2026
pdf
MS_Thesis_AmineDownloadView
Open Access

Abstract

Conformal Prediction Deep Learning Machine Learning Time Series Forecasting Uncertainty Quantification
Reliable uncertainty quantification is critical for streamflow forecasting, where predictive errors can have significant consequences for water resource management, flood mitigation, and drought planning. While modern deep learning models have achieved strong point prediction performance, their uncertainty estimates often lack formal guarantees and can degrade under temporal distribution shift. Conformal prediction provides a principled, distribution-free framework for constructing prediction intervals with finite-sample coverage guarantees, making it an attractive approach for hydrological forecasting. In this thesis, we investigate the application of conformal prediction methods to streamflow forecasting in a multi-basin setting. Using residuals from a deterministic deep learning forecasting model, we evaluate several residual-based conformal approaches, including pooled, volatility-scaled, autoregressive, and neural conditional quantile regression methods. We further adapt a relational conformal prediction framework to exploit temporal structure and cross-basin dependencies in residual dynamics. Through experiments on large-scale streamflow datasets, we analyze the coverage, sharpness, and robustness of different conformal methods under realistic temporal distribution shift. Our results highlight the challenges of uncertainty calibration in non-stationary hydrological time series and demonstrate that relational conformal prediction can produce substantially sharper prediction intervals while maintaining reliable coverage compared to residual-only baselines. These findings provide practical guidance for deploying conformal uncertainty quantification in real-world streamflow forecasting systems.

Metrics

1 Record Views

Details

Logo image