Medium- Long-Term Runoff Forecasting Using Interpretable Hybrid Machine Learning Model for Data-Scarce Regions
[Objectives] This study aims to analyze the applicability of existing precipitation, temperature, and runoff data in data-scarce regions, and to develop and evaluate a deep learning hybrid model driven by multi-source information for improving the accuracy of monthly runoff forecasting. [Methods] Ba...
Saved in:
| Main Author: | |
|---|---|
| Format: | Article |
| Language: | zho |
| Published: |
Editorial Office of Journal of Changjiang River Scientific Research Institute
2025-07-01
|
| Series: | 长江科学院院报 |
| Subjects: | |
| Online Access: | http://ckyyb.crsri.cn/fileup/1001-5485/PDF/1001-5485(2025)07-0052-08.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | [Objectives] This study aims to analyze the applicability of existing precipitation, temperature, and runoff data in data-scarce regions, and to develop and evaluate a deep learning hybrid model driven by multi-source information for improving the accuracy of monthly runoff forecasting. [Methods] Based on historical precipitation, temperature, and runoff sequences from the Yulongkashi River, a Convolutional Neural Network-Bidirectional Gated Recurrent Unit-Attention (CNN-BiGRU-Attention) model was developed. An Improved Particle Swarm Optimization (IPSO) algorithm was used to optimize this model, forming the IPSO-CNN-BiGRU-Attention hybrid model. The performance of this model was compared with that of the Gated Recurrent Unit (GRU) model and the ABCD water balance model. [Results] The IPSO-CNN-BiGRU-Attention model that incorporated precipitation and temperature data overall outperformed the CNN-BiGRU-Attention and GRU models, showing better agreement with the observed values. As the predication period increased, the proposed model achieved a root mean square error (RMSE) of 2.11 m3/s, a mean absolute error (MAE) of 1.32 m3/s, a mean absolute percentage error (MAPE) of 73.76%, and a Nash-Sutcliffe efficiency (NSE) coefficient of 0.94. The highest forecast accuracy was observed in the first three months. [Conclusions] The IPSO-CNN-BiGRU-Attention model effectively integrates precipitation, temperature, and runoff information, significantly enhancing the accuracy of monthly runoff forecasts in data-scarce regions. The model demonstrates robust performance across different forecast horizons, particularly suitable for short-term predictions of 1-3 months. This approach offers a practical and reliable tool for hydrological forecasting and flood control/drought management in data-scarce basins. |
|---|---|
| ISSN: | 1001-5485 |