Granata et al. (2026) Deep learning meets spectral analysis: reproducing long-term persistence in synthetic streamflow

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-09-29
• Authors: Francesco Granata, Fabio Di Nunno, Senlin Zhu
• DOI: 10.1016/j.jhydrol.2026.136519

Research Groups

• Department of Civil and Mechanical Engineering, University of Cassino and Southern Lazio, Via G. Di Biasio 43, 03043 Cassino, FR, Italy
• College of Hydraulic Science and Engineering, Yangzhou University, Yangzhou, China

Short Summary

This study evaluates a three-stage framework that combines latent-space generation with a temporal convolutional network (TCN), low-frequency spectral adjustment, and quantile-based rank reordering to reproduce long-term persistence in synthetic streamflow.

Objective

• Formulate a reproducible three-stage generator that separates neural simulation, spectral persistence adjustment, and marginal remapping.
• Quantify the framework's ability to reproduce marginal distributions, autocorrelation, persistence, seasonal behavior, and flow-spell characteristics across hydrologically diverse basins.

Study Configuration

• Spatial Scale: Daily discharge records from 24 UK river basins.
• Temporal Scale: Two-year windows with a 365-day overlap for training.

Methodology and Data

• Models used: Temporal Convolutional Network (TCN), ARMA, k-Nearest Neighbor (k-NN), Long Short-Term Memory (LSTM)
• Data sources: Satellite, observation, reanalysis

Main Results

• The framework reproduced marginal distributions closely (median Kolmogorov-Smirnov statistic = 0.00069; 99th-percentile ratios = 0.986–0.999).
• The median absolute difference in the rescaled-range persistence estimate was 0.028, with 79% of basins within a tolerance of 0.05.
• Detrended fluctuation analysis (DFA) produced a median scaling-exponent difference of 0.076.

Contributions

The proposed framework provides an explicit mechanism for targeting long-term persistence in synthetic streamflow generation, complementing established stochastic and learning-based approaches.

Funding

• This study was funded by the University of Cassino and Southern Lazio (Italy).
• Additional funding was provided by Yangzhou University (China).

Citation

@article{Granata2026Deep,
  author = {Granata, Francesco and Nunno, Fabio Di and Zhu, Senlin},
  title = {Deep learning meets spectral analysis: reproducing long-term persistence in synthetic streamflow},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136519},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136519}
}