Li et al. (2025) Multistep-ahead prediction of daily water temperature for Poyang Lake, China, using monthly monitoring data

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2025
• Date: 2025-12-11
• Authors: Gang Li, Xiting Li, Qixin Lin, Zhangjun Liu, Yuqin Gao, Zhen Cui
• DOI: 10.1016/j.ejrh.2025.103033

Research Groups

• Jiangxi Academy of Water Science and Engineering, Nanchang, China
• Jiangxi Key Laboratory of Flood and Drought Disaster Defense, Nanchang, China
• Key Laboratory of Poyang Lake Environment and Resources Utilization, Ministry of Education, Nanchang University, Nanchang, China
• College of Water Conservancy and Hydropower Engineering, Hohai University, Nanjing, China
• Nanjing University of Information Science & Technology, Nanjing, China

Short Summary

This study proposes a novel framework integrating a physically based (PB) model with deep learning (DL) models to address data scarcity for daily water temperature (WT) forecasting in large lakes. The framework successfully converts monthly WT observations into daily simulations and extends predictions to ungauged areas, demonstrating competitive multistep-ahead daily WT forecasts for Poyang Lake.

Objective

• To develop and validate a framework for predicting daily-scale water temperature (WT) dynamics using monthly monitoring data.
• To extend water temperature predictions from gauged stations to ungauged stations.

Study Configuration

• Spatial Scale: Poyang Lake (PYL), the largest freshwater lake in China. The lake was discretized into 50,866 grids, each 250 meters by 250 meters. The study covers both gauged and ungauged stations within the lake.
• Temporal Scale:
  • Data Resolution: Monthly water temperature observations; daily hydrological and atmospheric data; hourly water temperature data (averaged to daily) from one station.
  • Simulation Period: 2015–2016 (including a dry year and a wet year).
  • Prediction Horizons: One-day-ahead, three-day-ahead, and seven-day-ahead forecasts.

Methodology and Data

• Models used:
  • Physically Based (PB) Model: Environmental Fluid Dynamics Code (EFDC) model.
  • Deep Learning (DL) Models: Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model and Transformer model.
• Data sources:
  • Hydrological Boundary Data: Monthly data from 8 stations (QJ, WJB, WZ, LJD, MG, HS, DFK, HK) on inflowing and outflowing rivers.
  • Water Temperature (WT) Boundary Data: Monthly data from 10 stations (Xiuhekou, Gan1-4, LJD, MG, HS, DFK, HK) on inflowing and outflowing rivers.
  • Atmospheric Boundary Data: Daily data from XZ station within PYL (precipitation, evaporation, air temperature, sea level pressure, wind speed, wind direction).
  • In-lake WT Observations for Calibration/Validation:
    • Monthly: Hamashi (HMS), Xingzi (XZ), Duchang (DC), Kangshan (KS) for PB model calibration and temporal validation.
    • Monthly: Banghu (BH), Tangyin (TY) for independent spatial validation (regarded as ungauged).
    • Daily: Baisha (BS) station (April 1, 2015, to December 30, 2016) for high-frequency validation.
  • Topographic Data: 5 meter by 5 meter Digital Elevation Model (DEM) of PYL from 2010.

Main Results

• The physically based (PB) EFDC model successfully converted monthly water temperature (WT) observations into daily simulations, accurately capturing both long-term trends and intra-annual daily variations across the entire lake.
• The average Root Mean Squared Error (RMSE) for PB model simulated WT was 1.314 °C at calibrated stations and 1.950 °C at independent validation stations (BH, TY, BS).
• Deep learning (DL) models (CNN-LSTM and Transformer), trained with the PB-simulated daily WT, produced competitive multistep-ahead daily WT predictions at both gauged and ungauged stations.
• The CNN-LSTM model generally outperformed the Transformer model, showing lower error medians (MAE, RMSE, MAPE) and higher correlation coefficients (R), along with greater stability across multiple runs.
• For one-day-ahead forecasts (T+1), the minimum correlation coefficient (R) across all stations exceeded 0.97 when compared to observed WT, and 0.99 when compared to simulated WT. Maximum Mean Absolute Error (MAE) was less than 2.007 °C.
• For three-day-ahead (T+3) and seven-day-ahead (T+7) forecasts, the minimum R values remained above 0.92 (against observed WT) and 0.93 (against simulated WT), respectively, demonstrating satisfactory accuracy over longer horizons.
• The multivariate predictive strategy showed superior performance in fitting the simulated WT, particularly at shorter lead times, but the accuracy gap between multivariate and univariate strategies narrowed at longer lead times due to error propagation.

Contributions

• Proposed a novel integrated framework combining physically based (PB) and deep learning (DL) models to effectively address the challenge of data scarcity for daily water temperature (WT) prediction in large lakes.
• Developed a robust methodology to convert low-frequency (monthly) WT observations into high-frequency (daily) simulations and extrapolate these simulations to ungauged areas using a PB model.
• Demonstrated the successful application of DL models (CNN-LSTM and Transformer) for accurate multistep-ahead daily WT forecasting (1, 3, and 7 days) by leveraging PB-simulated daily data, achieving competitive performance even with limited original monthly observations.
• Provided a promising and practical solution for near-term daily WT prediction in data-scarce large lake environments, thereby maximizing the utility of existing monitoring data across diverse spatiotemporal scales.

Funding

• National Key Research and Development Program of China (2023YFC3206005)
• Open Research Foundation of Key Laboratory of Poyang Lake Environment and Resources Utilization, Ministry of Education (2022Z03)
• Science and Technology + Water Conservancy Joint Plan Project of Jiangxi Province (2023KSG01007)
• Science and Technology Project of Water Resources Department of Jiangxi Province (202425YBKT03, 202425YBKT02)
• Natural Science Foundation of Jiangxi Province (20242BAB20240)

Citation

@article{Li2025Multistepahead,
  author = {Li, Gang and Li, Xiting and Lin, Qixin and Liu, Zhangjun and Gao, Yuqin and Cui, Zhen},
  title = {Multistep-ahead prediction of daily water temperature for Poyang Lake, China, using monthly monitoring data},
  journal = {Journal of Hydrology Regional Studies},
  year = {2025},
  doi = {10.1016/j.ejrh.2025.103033},
  url = {https://doi.org/10.1016/j.ejrh.2025.103033}
}