MamGA: a deep neural network architecture for dual-channel parallel monthly runoff prediction based on mamba and depth-gated attention layer
•The first application of the Mamba architecture is to predict monthly runoff.•Development of a depth-gated attention layer to enhance bidirectional information capture.•Implement coding and decoding systems to capture temporal dynamics and enrich features.•Construction of a dual-channel parallel ar...
Uložené v:
| Vydané v: | Journal of hydrology (Amsterdam) Ročník 663; s. 134304 |
|---|---|
| Hlavní autori: | , , , |
| Médium: | Journal Article |
| Jazyk: | English |
| Vydavateľské údaje: |
Elsevier B.V
01.12.2025
|
| Predmet: | |
| ISSN: | 0022-1694 |
| On-line prístup: | Získať plný text |
| Tagy: |
Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
|
| Shrnutí: | •The first application of the Mamba architecture is to predict monthly runoff.•Development of a depth-gated attention layer to enhance bidirectional information capture.•Implement coding and decoding systems to capture temporal dynamics and enrich features.•Construction of a dual-channel parallel architecture (MamGA) for accurate runoff prediction.•MamGA achieves a balance between accuracy and efficiency, outperforming other comparison models.
Monthly runoff prediction is crucial in water resource management, involving both short-term hydrological dynamics and long-term planning. It has a decisive impact on flood prevention, resource allocation, and ecological protection. In the context of increasing uncertainties in runoff due to climate change and human activities, accurate monthly runoff forecasting becomes even more essential. Therefore, this paper proposes a novel dual-channel parallel monthly runoff prediction deep neural network architecture—MamGA—built on the significant application value of deep neural networks in runoff prediction. The architecture first introduces the Mamba model, which employs a selection mechanism to achieve selective information propagation and suppression, effectively enhancing the processing capability of global feature information while reducing the computational complexity of modelling long sequences. Furthermore, this paper incorporates a Depth-gated Attention Layer that combines bidirectional depth-gated modules and linear attention mechanisms to address the shortcomings of the Mamba network in unidirectional information processing. Integrating an Embedded Coding layer and a Sequential Decoding layer constructs an efficient coding and decoding system, further strengthening the model’s ability to capture global features and temporal information. To validate the effectiveness and advancement of the MamGA model, this study selected the Manwan Station (MW), Xiaowan Station (XW) in China, and the Thunder Creek Station (TC) in the United States as experimental subjects. Five evaluation metrics were employed for comparative analysis against nine benchmark models. The experimental results indicate that the MamGA model exhibits significant superiority across all cases. For instance, at the MW station, compared to the Long Short-Term Memory (LSTM) model, the MamGA model reduced the Mean Absolute Error (MAE) and Normalized Root Mean Square Error (NRMSE) by 33.08% and 23.93%, respectively. Meanwhile, the Nash Efficiency Coefficient (NSE), correlation coefficient (R), and Kling-Gupta Efficiency (KGE) improved by 8.41%, 3.93%, and 8.36%, respectively, with both R and NSE exceeding 0.9. The MamGA model also demonstrated significant performance improvements at other stations compared to the competing models. The study suggests that the MamGA model, as an advanced tool for monthly runoff prediction, can significantly enhance the accuracy of runoff forecasting, providing robust support for the optimal allocation and management of water resources. |
|---|---|
| ISSN: | 0022-1694 |
| DOI: | 10.1016/j.jhydrol.2025.134304 |