Scholarlink

<p>Reference evapotranspiration (ET<sub>0</sub>) is one of the core factors in hydrology, meteorology and agriculture. Accurate ET<sub>0</sub> prediction is vital for irrigation scheduling and water resource management. However, ET<sub>0</sub> prediction models typically rely solely on meteorological data while neglecting the potential information in historical ET<sub>0</sub> data, resulting in suboptimal prediction accuracy. In this paper, the seasonal-trend decomposition using Loess (STL) technique is adopted to gain the seasonal, trend and remainder components of the historical ET<sub>0</sub> data, which are integrated with meteorological data to enrich input information. An enhanced MHSA-BiLSTM model for ET<sub>0</sub> prediction is constructed by integrating the multi-head self-attention (MHSA) mechanism and bidirectional long short-term memory (BiLSTM) via a dual-modal parallel feature learning architecture. Daily meteorological datasets are collected from 3 weather stations in the Donglei Yellow River irrigation district. The predictions reveal that adding historical ET<sub>0</sub> features to the enhanced MHSA-BiLSTM model reduced its mean absolute error and mean square error by 78% and 60%, respectively, compared to using only meteorological data. The developed model shows a narrow error distribution and remarkable performance on all the datasets. The results demonstrate that the developed model outperforms MHSA-BiLSTM model, BiLSTM model, and MHSA model in prediction accuracy, with strong generalization ability across all datasets. The developed model enables highly accurate predictions of ET<sub>0</sub> over the next 10 days, with extremely precise results for the first 7 days.</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

Integrating Seasonal-Trend Decomposition and Multi-Head Self-Attention for Feature Fusion with Bidirectional Long Short-Term Memory in Reference Evapotranspiration Prediction

  • 期刊论文

  • 发表日期：

    2026年3月28日

查看全文
  • Zehai Gao,
  • Zijun Gao,
  • Chao Sun,
  • Xiaojun Zhang,
  • Siyuan Cao,
  • Yuwei Liu

摘要

Reference evapotranspiration (ET0) is one of the core factors in hydrology, meteorology and agriculture. Accurate ET0 prediction is vital for irrigation scheduling and water resource management. However, ET0 prediction models typically rely solely on meteorological data while neglecting the potential information in historical ET0 data, resulting in suboptimal prediction accuracy. In this paper, the seasonal-trend decomposition using Loess (STL) technique is adopted to gain the seasonal, trend and remainder components of the historical ET0 data, which are integrated with meteorological data to enrich input information. An enhanced MHSA-BiLSTM model for ET0 prediction is constructed by integrating the multi-head self-attention (MHSA) mechanism and bidirectional long short-term memory (BiLSTM) via a dual-modal parallel feature learning architecture. Daily meteorological datasets are collected from 3 weather stations in the Donglei Yellow River irrigation district. The predictions reveal that adding historical ET0 features to the enhanced MHSA-BiLSTM model reduced its mean absolute error and mean square error by 78% and 60%, respectively, compared to using only meteorological data. The developed model shows a narrow error distribution and remarkable performance on all the datasets. The results demonstrate that the developed model outperforms MHSA-BiLSTM model, BiLSTM model, and MHSA model in prediction accuracy, with strong generalization ability across all datasets. The developed model enables highly accurate predictions of ET0 over the next 10 days, with extremely precise results for the first 7 days.