Xu et al. (2025) The applicability of statistical post-processing techniques for quantitative precipitation forecast in the Huaihe River Basin

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2025
• Date: 2025-11-25
• Authors: Sunyu Xu, Ping-an Zhong, Xinyuan Qian, Bin Wang, Han Wang, Yiwen Wang, Weifeng Liu, Lixin Tian
• DOI: 10.1016/j.ejrh.2025.102988

Research Groups

• College of Hydrology and Water Resources, Hohai University, Nanjing, China.
• Changjiang Institute of Survey, Planning, Design and Research Co., LTD, Wuhan, China.
• MWR General Institute of Water Conservancy Resources and Hydropower Planning and Design, Beijing, China.
• China South-to-North Water Diversion Eastern Route Corporation Limited, Beijing, China.

Short Summary

This study evaluates seven post-processing methods for quantitative precipitation forecasts in the Huaihe River Basin, demonstrating that spatiotemporal deep learning models (specifically ConvLSTM) significantly outperform traditional statistical and time-series methods, particularly during flood seasons and in complex terrains.

Objective

• To evaluate the effectiveness and regional adaptability of various statistical, time-series, and spatiotemporal post-processing methods in improving quantitative precipitation forecasts (QPF) under diverse topographic, climatic, and seasonal conditions.

Study Configuration

• Spatial Scale: The upper Huaihe River Basin above Bengbu, China (approximately 121,000 km²), categorized into hilly vs. plain regions and humid vs. semi-humid zones.
• Temporal Scale: 15 years of data (2007–2021), analyzing daily precipitation with lead times ranging from 1 to 15 days, segmented into flood (June–September) and dry (October–May) seasons.

Methodology and Data

• Models used:
  • Statistical: Quantile Mapping (QM), Quantile Regression Forests (QRF).
  • Time Series: XGBoost, Long Short-Term Memory (LSTM), Transformer.
  • Spatiotemporal: Convolutional LSTM (ConvLSTM), Spatio-Temporal Graph Convolutional Network (STGCN).
• Data sources:
  • Forecasts: ECMWF ensemble mean forecasts (51 members) and meteorological predictors (temperature, wind, humidity, pressure) from the TIGGE database.
  • Observations: CN05.1 gridded daily precipitation dataset (0.25° × 0.25° resolution).
  • Topography: DEM-based elevation and topographic relief data.

Main Results

• Superiority of Spatiotemporal Modeling: ConvLSTM emerged as the best-performing model overall, achieving average improvement rates of 22.00% for RMSE, 46.75% for CC, and 15.56% for MAE compared to raw ECMWF forecasts.
• Lead Time Dynamics: QRF is highly effective for short-term forecasts (1–3 days), while STGCN and ConvLSTM show greater advantages for extended lead times (11–15 days).
• Environmental Adaptability:
  • Seasonality: All models perform better during the dry season due to stable precipitation; however, statistical models fail significantly during the flood season, where ConvLSTM maintains robust performance despite high variability.
  • Topography: Plain regions yield lower absolute errors (RMSE/MAE), but hilly regions show higher correlation coefficients (CC) as models better capture the stronger precipitation signals driven by orographic effects.
  • Climate: In humid zones with complex precipitation patterns, deep learning models provide more reliable corrections than traditional statistical methods.

Contributions

• Provides a systematic comparison of diverse post-processing categories (statistical vs. machine learning vs. deep learning) specifically for a transitional climate zone.
• Identifies the critical role of incorporating spatiotemporal features and environmental priors (topography and climate) to enhance regional forecast stability.
• Offers a practical selection framework for QPF post-processing based on lead time and seasonal requirements for disaster prevention and water resource management.

Funding

• National Key R&D Program of China (Grant No. 2022YFC3202801).
• Research project of China South-to-North Water Diversion Corporation Limited (DXZ-2023–081-ZD-ZX).

Citation

@article{Xu2025applicability,
  author = {Xu, Sunyu and Zhong, Ping-an and Qian, Xinyuan and Wang, Bin and Wang, Han and Wang, Yiwen and Liu, Weifeng and Tian, Lixin},
  title = {The applicability of statistical post-processing techniques for quantitative precipitation forecast in the Huaihe River Basin},
  journal = {Journal of Hydrology Regional Studies},
  year = {2025},
  doi = {10.1016/j.ejrh.2025.102988},
  url = {https://doi.org/10.1016/j.ejrh.2025.102988}
}

Generated by BiblioAssistant using gemini-3-flash-preview (Google API)