Author(s): Sait Mutlu Karahan; Wouter Vandenbruwaene; Jan Verwaeren

Linked Author(s):

Keywords: Electrical conductivity forecasting; Long Short-Term Memory; Uncertainty quantification

Abstract: Salinisation is a growing concern, requiring continuous, reliable water-quality monitoring to support effective environmental management. While data-driven models are increasingly used to forecast water quality parameters, their uncertainty quantification remains limited. In this study, an uncertainty-aware forecasting framework was developed using Long Short-Term Memory (LSTM) models to predict electrical conductivity (EC), a common proxy for salinisation. Upstream and downstream sensor locations along the Yser River (Belgium) were selected to capture spatial variability. One-step-ahead (15-minute) and multi-step-ahead (4hour) forecasting strategies were implemented to assess the effect of forecast horizon on predictive uncertainty. Aleatoric uncertainty was quantified using quantile regression, while epistemic uncertainty was estimated via Monte Carlo dropout. The results demonstrated clear spatial and temporal differences in predictive performance and in the characteristics of uncertainty estimation. For the upstream, one-step-ahead forecasting achieved low error metrics (RMSE = 0.008) and narrow uncertainty intervals (MPIW = 0.054), while multistep-ahead forecasts exhibited increased uncertainty (MPIW = 0.153). Downstream forecasts displayed wider uncertainty intervals, with MPIW increasing from 0.630 (one-step) to 1.236 (multi-step), reflecting environmental variability. Overall, uncertainty increased with forecast horizon, while reliable coverage was maintained across all configurations. The proposed framework provides a practical basis for uncertainty-aware interpretation of EC forecasts for river monitoring applications.

DOI: https://doi.org/10.64697/iahr.proc.hic2026.177

Year: 2026