摘要
针对现有气象业务中短临降雨预报不精确的问题,提出一种基于注意力机制的短临降雨预报方法。计算标准化降水指数后对因子作归一化处理;使用随机森林算法筛选出与降雨密切相关的气象因子;设计带有注意力机制的长短期记忆网络模型,有效解决长时间降雨序列信息丢失的问题,强化对关键信息的提取能力;对注意力权重可视化,提高模型透明度。在全国92个站点的实验结果验证了该方法有较好的泛化能力,在预测短临强降雨情况下比现有方法准确率更高。
Aimed at the problem of the inaccurate short-term rainfall forecast in meteorological operation,a short-term rainfall forecasting method based on attention mechanism is proposed.The factors were normalized after calculating SPI index.The meteorological factors closely related to rainfall were screened out by random forest algorithm.Long-short term memory network with attention mechanism was designed to effectively solve the information loss of long-time rainfall series and strengthen the ability to extract important information.In addition,attention weight was visualized to enhance transparency of the model.The experimental results at 92 stations in China show that the method has good generalization ability and has higher accuracy than the existing methods in predicting short-term and heavy rainfall.
作者
曹文南
张鹏程
贾旸旸
Cao Wennan;Zhang Pengcheng;Jia Yangyang(College of Computer and Information,Hohai University,Nanjing 210000,Jiangsu,China)
出处
《计算机应用与软件》
北大核心
2023年第2期48-53,81,共7页
Computer Applications and Software
基金
国家重点研发计划项目(2018YFC0407901)
河海大学中央高校基本科研业务费专项基金项目(2019B15414)。
