Scholarlink

<p>Accurate multi-step monthly rainfall forecasting is challenging because of rainfall non-stationarity and error accumulation at longer lead times. In this study, a hybrid Variational modal decomposition (VMD) - Long short time memory (LSTM)–cross-attention-mechanism (CATNet) framework that combines VMD-based signal regularization with a CATNet–enhanced Seq2Seq LSTM predictor to improve long-horizon stability is proposed. The model is evaluated using monthly meteorological data from three stations: arid inland (Ili Tianshan), coastal monsoon (Qingdao Liuting), and humid mountainous (Ganzhou) regions. The results show that the proposed framework consistently outperforms baseline methods for 2–6-month-ahead forecasting, achieving higher efficiency metrics in most cases and maintaining strong skill at longer lead times (e.g., Nash–Sutcliffe efficiency (NSE (6)) = 0.9682 at Ili Tianshan). Additional evaluations using the root mean square error (RMSE) and mean absolute error (MAE) further confirm the stability and accuracy of the proposed approach. These results indicate that the proposed framework provides robust and stable multi-step rainfall forecasting performance under contrasting climatic conditions in station-level case studies.</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

Multi-Feature and Multi-Step Monthly Rainfall Forecasting with the Hybrid VMD-LSTM-CATNet Model in Diverse Climatic Regions

  • 期刊论文

  • 发表日期：

    2026年3月19日

查看全文
  • Man-Hong Fan,
  • Xiao-Ying Ma,
  • Shi-Qi Xu,
  • Rui-Lin Fu,
  • Zhong-Xing Tao

摘要

Accurate multi-step monthly rainfall forecasting is challenging because of rainfall non-stationarity and error accumulation at longer lead times. In this study, a hybrid Variational modal decomposition (VMD) - Long short time memory (LSTM)–cross-attention-mechanism (CATNet) framework that combines VMD-based signal regularization with a CATNet–enhanced Seq2Seq LSTM predictor to improve long-horizon stability is proposed. The model is evaluated using monthly meteorological data from three stations: arid inland (Ili Tianshan), coastal monsoon (Qingdao Liuting), and humid mountainous (Ganzhou) regions. The results show that the proposed framework consistently outperforms baseline methods for 2–6-month-ahead forecasting, achieving higher efficiency metrics in most cases and maintaining strong skill at longer lead times (e.g., Nash–Sutcliffe efficiency (NSE (6)) = 0.9682 at Ili Tianshan). Additional evaluations using the root mean square error (RMSE) and mean absolute error (MAE) further confirm the stability and accuracy of the proposed approach. These results indicate that the proposed framework provides robust and stable multi-step rainfall forecasting performance under contrasting climatic conditions in station-level case studies.