Related Experiment Video
Updated: Jul 13, 2026
14:08
Automated Midline Shift and Intracranial Pressure Estimation based on Brain CT Images
Published on: April 13, 2013
42.5K
Multi-step ahead forecasting of daily streamflow based on the transform-based deep learning model under different
Miao He1,2, Xian Xu1, Shaofei Wu3
1Jiangxi Provincial Key Laboratory of Water Resources Allocation and Efficient Utilization, Nanchang Institute of Technology, Nanchang, 330099, China.
Scientific Reports
|February 14, 2025
Summary
The Rel-Informer model shows superior performance in short-term runoff prediction compared to LSTM, Transformer, and standard Informer models. It also effectively predicts runoff in ungauged catchments, offering a promising tool for hydrological forecasting.
Area of Science:
- Hydrology
- Data Science
- Machine Learning
Background:
- Accurate runoff prediction is crucial for water resource management and flood control.
- Deep learning models like LSTM have advanced runoff prediction, but newer models like Transformer and Informer show potential for improvement.
- Research on multi-step runoff prediction using advanced deep learning models across diverse scenarios is limited.
Purpose of the Study:
- To introduce and evaluate a novel relative location coding-enhanced Informer model (Rel-Informer) for rainfall-runoff prediction.
- To compare Rel-Informer's performance against standard Informer, Transformer, and LSTM models across various modeling scenarios.
- To assess the model's effectiveness in individual, regional, and large-scale (ungauged catchment) runoff prediction.
Main Methods:
- Utilized the publicly available CAMELS dataset for model training and validation.
- Implemented four experimental designs: individual, regional, fine-tuned regional, and large-scale modeling for ungauged catchments.
- Compared the performance of Rel-Informer, standard Informer, Transformer, and LSTM models using rainfall and runoff data.
Main Results:
- Rel-Informer consistently outperformed other models, especially in short-term (1-3 days) runoff predictions.
- Fine-tuning significantly improved the precision of regional rainfall-runoff models.
- The large-scale regional Rel-Informer model demonstrated effective runoff prediction for ungauged catchments.
- Hydrological characteristics like snowmelt and baseflow indices were found to influence prediction accuracy.
Conclusions:
- The Rel-Informer model, with enhanced relative position encoding, is a highly promising tool for runoff forecasting.
- The model shows particular strength in data-rich catchments and for short-term prediction tasks.
- Rel-Informer's capability in large-scale regional modeling offers a viable solution for predicting runoff in ungauged basins.
