Jia et al. (2026) Enhancing Streamflow Prediction Using Cutting-edge Deep Learning Models and Seasonal-Trend Decomposition

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-02-01
• Authors: Yubo Jia, Xiaoling SU, H. Felix Wu, Hanwen Yan, Xinxing Zhu
• DOI: 10.1007/s11269-025-04447-5

Research Groups

• Key Laboratory of Agricultural Soil and Water Engineering in Arid and Semiarid Areas, Ministry of Education, Northwest A & F University, Yangling, China
• College of Water Resources and Architectural Engineering, Northwest A & F University, Yangling, Shaanxi, China

Short Summary

This study systematically evaluates three cutting-edge deep learning models (FITS, FGN, PatchTST) against traditional models (LSTM, CNN, GRU) for streamflow prediction, both with and without four seasonal-trend decomposition (STD) techniques. It demonstrates that advanced models, particularly FITS, offer superior accuracy, robustness, and computational efficiency, while STD significantly improves traditional models but has limited impact on the advanced ones.

Objective

• To systematically evaluate the performance (accuracy, robustness, and computational efficiency) of three cutting-edge deep learning models (FITS, FGN, PatchTST) against traditional deep learning models (LSTM, CNN, GRU) for streamflow forecasting.
• To assess the impact of four advanced seasonal-trend decomposition (STD) techniques (MOV, LD, EXP, DFT-MOV) on the predictive performance of both traditional and advanced deep learning models, providing guidance for optimal data preprocessing strategies.

Study Configuration

• Spatial Scale: Jialing River basin, China, focusing on four representative hydrological stations: Fengzhou (Fz), Lueyang (Ly), Tingzikou (Tzk), and Beibei (Bb). The basin area is 159,800 square kilometers.
• Temporal Scale: Daily streamflow data from 2 January 2010 to 30 November 2022.

Methodology and Data

• Models used:
  • Deep Learning Models: Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU) (traditional baselines); FITS (Multi-Layer Perceptron architecture), FGN (Graph Convolutional Network architecture), PatchTST (Transformer architecture) (cutting-edge models).
  • Seasonal-Trend Decomposition (STD) Techniques: Moving Average Kernel (MOV), Learnable Decomposition (LD), Exponential Moving Average (EXP), Discrete Fourier Transform–Moving Average kernel (DFT-MOV).
  • Hybrid Models: 24 combinations formed by pairing each of the six deep learning models with each of the four STD techniques. The methodology involved sequence decomposition, component prediction (seasonal by DL model, trend by single-layer Multilayer Perceptron (MLP)), and result reconstruction.
  • Feature Selection: Least Absolute Shrinkage and Selection Operator (LASSO) regression was used to select the top eight most relevant hydrometeorological variables.
• Data sources:
  • Streamflow data: Provided by the Bureau of Hydrology, Yangtze Water Resources Commission, Ministry of Water Resources of China.
  • Hydrometeorological factors (11 variables) and upstream flow data: Obtained from the China Meteorological Data Service Centre of the National Meteorological Information Center, China Meteorological Administration (https://data.cma.cn/).
  • Data Preprocessing: Z-score scaling for input variables. Dataset split into training, validation, and test sets in a 6:2:2 ratio.
  • Hyperparameter Optimization: Bayesian optimization was employed for efficient hyperparameter search.
  • Computational Environment: Intel Xeon Platinum 8362 2.80 GHz CPU, 45 GB RAM, NVIDIA GeForce RTX 3090 24 GB GPU, PyTorch framework.

Main Results

• Performance of Standalone Deep Learning Models:
  • Cutting-edge models (FITS, FGN, PatchTST) significantly outperformed traditional models (LSTM, CNN, GRU) in accuracy and robustness across all four stations.
  • FITS achieved the highest predictive accuracy (mean Nash–Sutcliffe efficiency, NSE = 0.986; mean Kling–Gupta efficiency, KGE = 0.986) and exhibited superior stability (smallest standard deviation across performance metrics).
  • At Fz station, FITS improved NSE by 2.6%, KGE by 16.9%, and reduced MAPE by 1.5% and RMSE from 10.6 cubic meters per second to 7.7 cubic meters per second compared to the best traditional model (GRU).
  • Prediction accuracy was generally higher at downstream stations (Tzk, Bb) than at upstream stations (Ly, Fz).
• Impact of STD Techniques on Model Performance:
  • STD techniques substantially improved the performance of traditional models (LSTM, CNN, GRU). For CNN, NSE increased by 0.65%–1.76%, KGE by 3.92%–4.90%, RMSE decreased by 13.8%–35.6%, and MAPE decreased by 13.4%–37.7%.
  • The performance gains for advanced models (FITS, FGN, PatchTST) from STD techniques were limited (mean NSE changes of -0.13%–0.52%).
  • The hybrid model combining FGN with DFT-MOV decomposition showed the best overall performance among hybrid models.
  • Different STD techniques exhibited specific optimization effects for different models (e.g., EXP for LSTM, LD for CNN, DFT-MOV for GRU and FGN).
• Extreme Flow Prediction:
  • STD techniques enhanced the predictive capability for extreme streamflow events. MOV, EXP, and DFT-MOV significantly reduced high-flow prediction bias for traditional models.
  • MOV, LD, and DFT-MOV effectively reduced low-flow prediction biases across all benchmark models.
• Computational Efficiency:
  • FITS demonstrated the highest computational efficiency among standalone models, with a training time of 0.19 seconds and an inference time of 0.03 seconds (at Fz station), significantly lower than other models.
  • Among STD techniques, MOV exhibited the best computational efficiency, with training time increases of 13.9%–42.8% and inference time increases of 25.2%–55.4% relative to baseline models.
  • For traditional models, integrating STD techniques achieved a favorable balance between accuracy gains and computational cost, with training and inference time increases kept below 100% for substantial error reduction.

Contributions

• First-time introduction and systematic benchmarking of three cutting-edge deep learning architectures (FITS, FGN, PatchTST) for streamflow forecasting within the hydrological domain.
• Comprehensive evaluation of four advanced seasonal-trend decomposition (STD) techniques (MOV, LD, EXP, DFT-MOV) and their differential impacts on both traditional and advanced deep learning models.
• Provision of valuable insights and practical guidance for selecting and optimizing STD techniques and streamflow forecasting models, particularly highlighting the superior performance and efficiency of FITS.
• A systematic multi-model comparative framework that clarifies the strengths and weaknesses of various models in terms of accuracy, robustness, extreme-flow reproduction, and computational efficiency.

Funding

• Key Science and Technology Program of Shaanxi Province (2024SLKJ-14)

Citation

@article{Jia2026Enhancing,
  author = {Jia, Yubo and SU, Xiaoling and Wu, H. Felix and Yan, Hanwen and Zhu, Xinxing},
  title = {Enhancing Streamflow Prediction Using Cutting-edge Deep Learning Models and Seasonal-Trend Decomposition},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-025-04447-5},
  url = {https://doi.org/10.1007/s11269-025-04447-5}
}