Scholarlink

<p> <?tk 1?>Accurate precipitation prediction has been a major challenge in hydrological research. This study develops a novel model called GRDE by combining the Random Distribution Embedding (RDE) and Gated Recurrent Unit (GRU). The GRDE integrates spatiotemporal information from high-dimensional data, overcoming constraints from limited samples. Firstly, the data's dimension is increased by using phase space reconstruction (PSR), and the original high-dimensional attractor of the nonlinear dynamic system is reconstructed. Secondly, the RDE model is applied to generate a large amount of non-delay attractors as well as delay attractor of the target variable. Finally, the GRU model is used to create a mapping between the non-delay attractors and the delay attractor, converting the spatial information inherent in high-dimensional data into future information pertaining to the target variable. The GRDE model is applied to predict the monthly precipitation anomaly percentage at 17 meteorological stations located in the middle and lower sections of the Yangtze River, China. The result shows that the Temporal Correlation Coefficient (TCC), the mean anomaly correlation coefficient (ACC), and the average Percentage Coincident Rate (PCR) and the average trend anomaly comprehensive score (PS) are much higher than those of Temporal Convolutional Network (TCN), the Sequence-to-Sequence model (Seq2Seq), GRU and RDE; the ACC of GRDE is 45.9% to 443.7% higher than that of TCN, Seq2Seq, GRU and RDE. Moreover, the GRDE model shows the highest level of precision in predicting extreme precipitation, displaying a robust association between the predicted and observed value. Thus, the GRDE exhibits higher predictive capabilities for the spatial distribution, positive and negative trends, and anomaly levels of precipitation in the middle and lower portions of the Yangtze River.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A coupled model of nonlinear dynamical and deep learning for monthly precipitation prediction with small samples

  • 期刊论文

  • 发表日期：

    2025年3月8日

查看全文
  • Lili Deng,
  • Longxia Qian,
  • Mei Hong,
  • Dongyu Li,
  • Yijia Hu

摘要

Accurate precipitation prediction has been a major challenge in hydrological research. This study develops a novel model called GRDE by combining the Random Distribution Embedding (RDE) and Gated Recurrent Unit (GRU). The GRDE integrates spatiotemporal information from high-dimensional data, overcoming constraints from limited samples. Firstly, the data's dimension is increased by using phase space reconstruction (PSR), and the original high-dimensional attractor of the nonlinear dynamic system is reconstructed. Secondly, the RDE model is applied to generate a large amount of non-delay attractors as well as delay attractor of the target variable. Finally, the GRU model is used to create a mapping between the non-delay attractors and the delay attractor, converting the spatial information inherent in high-dimensional data into future information pertaining to the target variable. The GRDE model is applied to predict the monthly precipitation anomaly percentage at 17 meteorological stations located in the middle and lower sections of the Yangtze River, China. The result shows that the Temporal Correlation Coefficient (TCC), the mean anomaly correlation coefficient (ACC), and the average Percentage Coincident Rate (PCR) and the average trend anomaly comprehensive score (PS) are much higher than those of Temporal Convolutional Network (TCN), the Sequence-to-Sequence model (Seq2Seq), GRU and RDE; the ACC of GRDE is 45.9% to 443.7% higher than that of TCN, Seq2Seq, GRU and RDE. Moreover, the GRDE model shows the highest level of precision in predicting extreme precipitation, displaying a robust association between the predicted and observed value. Thus, the GRDE exhibits higher predictive capabilities for the spatial distribution, positive and negative trends, and anomaly levels of precipitation in the middle and lower portions of the Yangtze River.