Jiang et al. (2026) A parallel attention-based framework for multi-step multivariate runoff forecasting in mountainous watersheds: Wuyuan case study

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-01-06
• Authors: Jiange Jiang, Chen Chen, Yang Zhou, Wei Han, Lei Liu, Qingqi Pei
• DOI: 10.1016/j.ejrh.2025.103045

Research Groups

• School of Information Engineering, Jingdezhen Ceramic University, China
• School of Telecommunications Engineering, Xidian University, China
• The Ministry of Water Resources of China, China
• Xi’an Beilin University-based Innovation Group Co., Ltd., China
• Xidian Hangzhou Institute of Technology, Xidian University, China

Short Summary

This paper proposes a Parallel Attention Multivariate Multi-step LSTM Network (PAM2-LSTM) to address challenges in multivariate modeling and error accumulation in multi-step runoff forecasting. The model significantly outperforms conventional methods, achieving a 70.5% improvement in Mean Absolute Error (MAE) and a 61.2% reduction in Root Mean Square Error (RMSE) for 6-hour ahead forecasts, maintaining robust accuracy across extended prediction horizons.

Objective

• To develop a deep learning framework that explicitly captures heterogeneous relationships between hydrological variables and mitigates error propagation in multi-step ahead runoff forecasting for mountainous watersheds.

Study Configuration

• Spatial Scale: Wuyuan County, Jiangxi Province, China (2947 km²), characterized by mountainous terrain and complex river networks. Also evaluated on Xixian watershed, Henan Province, China (10,190 km²), with flatter topography, to demonstrate generalizability.
• Temporal Scale: Five-year period from 1 January 2017 to 4 December 2021, with hourly measurements (43,161 records). Forecast horizons evaluated at 1 hour, 2 hours, 6 hours, and 12 hours ahead.

Methodology and Data

• Models used: Parallel Attention Multivariate Multi-step LSTM Network (PAM2-LSTM).
  • Architecture: Encoder-decoder with parallel subnetworks for distinct variable groups (e.g., meteorological, hydrological state variables).
  • Forecasting strategy: Direct multi-step forecasting (dedicated predictor for each lead time) to mitigate error accumulation.
  • Enhancement: Attention mechanism to dynamically weight historical inputs based on relevance.
  • Baseline models for comparison: LSTM, GRU, BiLSTM, BiGRU, Seq2Seq.
• Data sources:
  • High-resolution observational data from Wuyuan watershed, Jiangxi Province, China, and Xixian watershed, Henan Province, China.
  • Wuyuan dataset: Hourly water level and discharge from Sandu hydrological station, complemented by precipitation and temperature data from 35 meteorological stations.
  • Data pre-processing: Quality control (temporal consistency, outlier detection), gap filling (spatial interpolation), linear/kriging interpolation to hourly intervals, Min-Max normalization.
  • Feature selection: Autocorrelation coefficients determined a sliding window length of 302 hours for runoff. Spearman correlation coefficients (threshold 𝜌 ≥ 0.4) selected historical input durations of 13 hours for rainfall, 51 hours for reservoir level, and 57 hours for water level.

Main Results

• The PAM2-LSTM model significantly outperforms conventional serial architectures and other baseline models in runoff forecasting.
• For 6-hour ahead forecasts in the Wuyuan watershed, PAM2-LSTM achieved a 70.5% improvement in MAE and a 61.2% reduction in RMSE compared to the conventional LSTM.
• Nash–Sutcliffe Efficiency (NSE) values exceeded 0.97 for 6-hour forecasts, indicating excellent model performance.
• The model maintains robust accuracy across extended prediction horizons, with a sub-10% relative error at 12-hour ahead forecasts (8.79% average relative error, 87.8% reduction in runoff error compared to iterative LSTM).
• Visualization of hydrographs confirms precise temporal alignment with observed data and reliable peak flow estimation, crucial for flash flood warning systems.
• The parallel architecture effectively leverages multivariate data by capturing heterogeneous relationships, while the direct multi-step forecasting strategy and attention mechanism minimize cumulative errors.
• The model's generalizability was confirmed through successful application to the Xixian watershed, demonstrating consistent superior performance.

Contributions

• Proposed a novel parallel architecture that explicitly models heterogeneous relationships between runoff and various hydrological variables, enabling more effective utilization of multivariate data and supporting future integration of remote sensing data.
• Developed an attention-based encoder-decoder framework that combines direct forecasting with iterative refinement, significantly reducing error accumulation in multi-step ahead predictions.
• Validated through extensive experiments on diverse watershed datasets that PAM2-LSTM outperforms existing methods in both accuracy and reliability for operational flood forecasting scenarios.

Funding

• National Natural Science Foundation of China (62072360, 62172438)
• Key research and development plan of Shaanxi province (2021ZDLGY02-09, 2023-GHZD-44, 2023-ZDLGY-54)
• National Key Laboratory Foundation (2023-JCJQ-LB-007)
• Natural Science Foundation of Guangdong Province of China (2022A1515010988)
• Key Project on Artificial Intelligence of Xi’an Science and Technology Plan (23ZDCYJSGG0021-2022, 23ZDCYYYCJ0008, 23ZDCYJSGG0002-2023)
• Xidian-UTAR China Malaysia Science and Technology Institute-the Fundamental Research Funds for the Central Universities/XURF-2024-QTZX24089
• Early-Career Young Scientists and Technologists Project of Jiangxi Province (S202510280)
• Guidance Plan Project for Industrial Development of Jingdezhen (2025GY019)
• Proof-of-concept fund from Hangzhou Research Institute of Xidian University (GNYZ2023QC0201, GNYZ2024QC004, GNYZ2024QC015)

Citation

@article{Jiang2026parallel,
  author = {Jiang, Jiange and Chen, Chen and Zhou, Yang and Han, Wei and Liu, Lei and Pei, Qingqi},
  title = {A parallel attention-based framework for multi-step multivariate runoff forecasting in mountainous watersheds: Wuyuan case study},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2025.103045},
  url = {https://doi.org/10.1016/j.ejrh.2025.103045}
}