Privacy-Preserving Federated Learning for Hydrological Forecasting in the Chu–Talas Basin

Raushan Amanzholova, Azamat Serek, Adil Akhmetov, Zhuldyzbek Onglassynov, S. A. Kulbekova, И.К. Рахметов, Jay Sagin

2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.3390/w18111361
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

The hydrological prediction in transboundary river basins is difficult because of their heterogeneous data distribution, restrictions of data sovereignty, and the irregular nature of the discharge pattern. In this paper, Federated Learning (FL) with an LSTM neural network is proposed for next-day predictions of discharge in the Chu–Talas transboundary basin. The basin area belongs to both Kazakhstan and Kyrgyzstan. In an FL scenario, two hydrological stations from the basin were selected as client nodes, representing two different discharge regimes. Station 15125—the Chu main channel is characterized by the highest discharge regime among stations located in the basin, while Station 15233—Merke tributary represents a small catchment with an irregular regime. The federated LSTM model is compared against a centralized LSTM and a local-only LSTM baseline model. The training process is based on nearly three decades of daily discharge measurements. The preprocessing step includes synchronization, lag calculation, and windowing operation. The models are trained using three metrics: root mean square error, mean absolute error, and Nash–Sutcliffe Efficiency, as well as using Monte Carlo Dropout for estimation of the probabilistic uncertainty. The results demonstrate that the federated model demonstrates comparable performance with the centralized one for the Chu main channel. It also improves prediction accuracy for the smaller Merke tributary compared with both centralized and local-only models. These findings show that FL can work effectively with non-IID and heterogeneous hydrological data. The study makes three main contributions: (i) it implements the FedAvg algorithm on transboundary, heterogeneous hydrological data, proving that decentralized optimization can effectively capture autoregressive temporal hydrology without data centralization; (ii) it systematically compares federated, centralized, and local-only models, demonstrating that the federated approach eliminates the scale bias that traditionally neglects smaller, high-variance catchments; and (iii) it utilizes Monte Carlo Dropout to translate deterministic AI outputs into risk-aware probabilistic bounds. Ultimately, the results of this study demonstrate the practical and scientific usefulness of FL in operational water management, as the method presents a privacy-saving means of increasing predictive capacity and enabling risk-based decision-making in transboundary river basins.