Scholarlink

<p>Forecasting seasonal rainfall is challenging but has important socio-economic implications. In Vietnam, despite encouraging results from previous studies, practical application remains difficult. This is because forecast accuracy remains low when studies only evaluate and use the products of global or regional models, or correct forecasts using simplistic traditional statistical methods. Modern statistical models using artificial intelligence show great potential to improve seasonal rainfall forecasting. This study aims to build a highly accurate forecasting tool for rainfall from May to October in seven climatic regions of Vietnam using a hybrid approach that combines convolutional neural networks (CNNs) and the dynamic downscaling product of the climate Weather Research and Forecasting model (clWRF). Three CNN models were developed for lead times of 1, 3, and 5&#xa0;months. To train the models, the study used monthly data from 1983 to 2011 (29&#xa0;years), including 5 meteorological variables from clWRF forecasts and monthly rainfall observations. The forecasts from the CNN models were then evaluated against observations using the dataset for 2012–2020 (9&#xa0;years). The results showed that the distribution of rainfall forecasted using CNN models closely agreed with the observations. In addition, the performance of the CNN models was supported by low Relative Mean Absolute Error (mostly below 30% for the regional average and 50% at each grid point), along with reasonably high spatial correlations with observed patterns (0.4–0.8). These results outperformed the forecasts produced by the clWRF model, with Added Value mostly positive, ranging from 0.2 to 0.8. This study contributes to enhancing the practical application of seasonal rainfall forecast information.</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 Hybrid Approach for Seasonal Rainfall Forecasting Across Vietnam Using Convolutional Neural Networks and Dynamical Downscaling

  • 期刊论文

  • 发表日期：

    2025年10月8日

查看全文
  • Thanh Pham Ngoc,
  • Huu Duy Nguyen,
  • Long Trinh-Tuan,
  • Phu Nguyen Duc,
  • Toan Trinh,
  • Tan Phan-Van

摘要

Forecasting seasonal rainfall is challenging but has important socio-economic implications. In Vietnam, despite encouraging results from previous studies, practical application remains difficult. This is because forecast accuracy remains low when studies only evaluate and use the products of global or regional models, or correct forecasts using simplistic traditional statistical methods. Modern statistical models using artificial intelligence show great potential to improve seasonal rainfall forecasting. This study aims to build a highly accurate forecasting tool for rainfall from May to October in seven climatic regions of Vietnam using a hybrid approach that combines convolutional neural networks (CNNs) and the dynamic downscaling product of the climate Weather Research and Forecasting model (clWRF). Three CNN models were developed for lead times of 1, 3, and 5 months. To train the models, the study used monthly data from 1983 to 2011 (29 years), including 5 meteorological variables from clWRF forecasts and monthly rainfall observations. The forecasts from the CNN models were then evaluated against observations using the dataset for 2012–2020 (9 years). The results showed that the distribution of rainfall forecasted using CNN models closely agreed with the observations. In addition, the performance of the CNN models was supported by low Relative Mean Absolute Error (mostly below 30% for the regional average and 50% at each grid point), along with reasonably high spatial correlations with observed patterns (0.4–0.8). These results outperformed the forecasts produced by the clWRF model, with Added Value mostly positive, ranging from 0.2 to 0.8. This study contributes to enhancing the practical application of seasonal rainfall forecast information.