DTTR: Encoding and decoding monthly runoff prediction model based on deep temporal attention convolution and multimodal fusion

•A DTTR deep learning model is proposed for accurate forecasting of monthly runoff.•Enhancing Feature Acquisition with Deep Convolutional Residual Networks (DCRN).•Applying temporal attention for dynamic feature weighting in DCRN outputs.•Improved inter-global feature feedback via encoder-decoder ar...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Journal of hydrology (Amsterdam) Ročník 643; s. 131996
Hlavní autoři: Wang, Wen-chuan, Tian, Wei-can, Hu, Xiao-xue, Hong, Yang-hao, Chai, Fu-xin, Xu, Dong-mei
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier B.V 01.11.2024
Témata:
Attention mechanism
climate change
Coding and decoding architecture
Convolutional residual network
Deep learning model
humans
model validation
Monthly runoff forecasting
prediction
rivers
runoff
water management
Attention mechanism
Monthly runoff forecasting
Deep learning model
Convolutional residual network
Coding and decoding architecture
ISSN:0022-1694
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:•A DTTR deep learning model is proposed for accurate forecasting of monthly runoff.•Enhancing Feature Acquisition with Deep Convolutional Residual Networks (DCRN).•Applying temporal attention for dynamic feature weighting in DCRN outputs.•Improved inter-global feature feedback via encoder-decoder architecture. Accurate runoff forecasting facilitates effective water resource management, and ensures the sustainable allocation of water for agricultural, industrial, and domestic use. Accurate runoff prediction has become more challenging due to the increased complexity associated with climate change and human activities. This paper proposes a new forecasting model, Deep Convolutional Residual Network with Temporal Attention and Transformer (DTTR), which is innovatively embedded with a temporal attention deep convolutional network to form a multimodal fusion “encoding-decoding” architecture. First, the weight allocation of higher-order hidden features extracted by the Deep Convolutional Residual Network (DCRN) is optimized by introducing the Temporal Attention Mechanism (TAM) to enhance the capture ability of sequence features. Secondly, the model adopts the “encoding–decoding” architecture to extend the feature dimensions and learns the temporal location information to enhance the global feature inter-feeding. Finally, the DTTR model successfully integrates the global information and maps the sequence features from multiple perspectives, which significantly improves the data’s feature abstraction ability and thus realizes the accurate prediction of the monthly runoff sequence. To verify the validity and sophistication of the DTTR model, the Taolai River, Hongshan River, and Fengle River were selected as experimental subjects. The model performance was tested using five evaluation indicators and nine comparison models. The results show that the DTTR model performs better than the benchmark model in different cases. For example, at Fengle River station, the mean absolute error (MAE), normalized root mean square error (NRMSE), Nash-Sutcliffe efficiency coefficient (NSE), correlation coefficient (R), and Kling-Gupta efficiency (KGE) metrics of the DTTR model are improved by 30.49%, 37.18%, 7.87%, compared to the LSTM model, 3.16% and 10.34%. The R and KGE of each site exceeded 0.9, and the DTTR model also showed significant performance improvement in other cases. The experimental results demonstrate that the DTTR model, as an advanced model for predicting menstrual flow, can help to improve the accuracy of monthly runoff prediction and support the subsequent development of water resource optimization allocation and management plans.
Bibliografie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 23
ISSN:0022-1694
DOI:10.1016/j.jhydrol.2024.131996