Enhancement of hydrological time series prediction with Real-World Time Series Generative Adversarial Network-based synthetic data and deep learning models

Tools

Dodig, Ana and Stankovic, Vladimir and Stankovic, Lina and Stojkovic, Milan (2026) Enhancement of hydrological time series prediction with Real-World Time Series Generative Adversarial Network-based synthetic data and deep learning models. Environmental Modelling and Software, 204. 107037. ISSN 1364-8152 (https://doi.org/10.1016/j.envsoft.2026.107037)

[thumbnail of Dodig-etal-EMS-2026-Enhancement-of-hydrological-time-series-prediction]
Preview
 Text. Filename: Dodig-etal-EMS-2026-Enhancement-of-hydrological-time-series-prediction.pdf
Accepted Author Manuscript
License: Creative Commons Attribution 4.0 logo
Download (14MB)| Preview

Abstract

Accurate river water quality forecasting is essential for environmental and public health. However it is challenging due to limited high-resolution data. To overcome this, a comprehensive framework for data augmentation and forecasting is developed. The proposed approach leverages available flow and water temperature (WT) measurements alongside sparse dissolved oxygen (DO) observations to enhance the temporal resolution of the dataset and employ Real-World Time Series Generative Adversarial Network (RTSGAN) to create additional synthetic dataset that retain real-world data characteristics. Both the original and synthetic data are then used to train two deep learning models, Graph Neural Networks (GNNs) and Long Short-Term Memory networks (LSTMs), which predict DO and WT at three hydrological stations three days ahead. The results indicate that GNNs, through the combined use of synthetic data and spatial dependencies, achieve up to 43.4% higher accuracy at upstream locations, whereas LSTMs improve predictive performance by up to 15%.

ORCID iDs

Dodig, Ana, Stankovic, Vladimir ORCID logoORCID: https://orcid.org/0000-0002-1075-2420, Stankovic, Lina ORCID logoORCID: https://orcid.org/0000-0002-8112-1976 and Stojkovic, Milan;
CORE (COnnecting REpositories)