摘要
遵循AI技术典型应用流程,从确定研究对象的数据表示开始,探讨从测试选用算法工具构建“区域性暴雨”深度学习预测模型,到调整优化模型超参数、对模型性能进行泛化增强,最终实现上线部署,成功对2023年广西“龙舟水”过程作出合理预测。结果表明,结合实际预报服务需求,在深入理解算法和模型架构特点基础上,选用能够模拟天气分析过程、合理解释预测机理的AI算法,可显著增强模型的实用性能。采用TimeDistributed层对样本时间层进行封装,先提取学习气象要素场空间特征和关联特征后再进一步学习时间变化知识,是一种符合天气分析思路且预测性能较好的气象AI模型构建方法。而针对“小概率”气象关注事件,在设定事件标签时适当降低气象评判指标、基于相似天气形势来增强样本数据,引导模型针对关注目标进行正向增强训练,是提升气象AI模型性能的有效技术手段,实现一个“有/无区域性暴雨发生”定性预报的基础分类模型。
Following a typical application process of AI technology,and starting from determining the data representation of the research object,this article discusses the construction process of a deep learning prediction model for“regional rainstorm”from using testing and selecting algorithmic tools,to adjusting and optimizing the model hyperparameters,enhancing the generalization of the model's performance,and ultimately achieving its deployment online,thus successfully making sensible predictions of the“Dragon-boat Rain”event in Guangxi in 2023.The results show that the practical performance of the model can be significantly enhanced by employing AI algorithms that can simulate the weather analysis process and provide reasonable explanations for prediction mechanisms based on an in-depth understanding of the characteristics of the algorithm and model architecture in combination with the actual demand for forecasting services.The application of TimeDistributed layers to encapsulate the temporal dimension of samples,which first extracts the spatial and related features of meteorological elements before further learning about temporal changes,is a meteorological AI model construction method that aligns with weather analysis approaches and shows better predictive performance.For“low probability”meteorological events of interest,it is an effective technical means to improve the performance of weather AI models by appropriately lowering meteorological evaluation indicators when setting event labels,enhancing sample data based on similar weather situations,and guiding the model to carry out the positive enhancement training for the focus of attention,so as to achieve a basic classification model for qualitative prediction of“with/without regional rainstorm occurrence”.
作者
覃月凤
董良淼
刘国忠
梁存桂
杨明鑫
QIN Yuefeng;DONG Liangmiao;LIU Guozhong;LIANG Cungui;YANG Mingxin(Guangxi Meteorological Observatory,Nanning 530022,China)
出处
《气象研究与应用》
2024年第3期29-36,共8页
Journal of Meteorological Research and Application
基金
中国气象局“揭榜挂帅”项目(CMAJBGS202217)
中国气象局气象能力提升联合研究专项(22NLTSY011)
广西气象科研计划项目(桂气科2024M03、桂气科2019M07)。
作者简介
第一作者:覃月凤(1988-),工程师,主要从事天气预报技术研究。E-mail:qinyfjune@163.com;通讯作者:董良淼(1973-),正高级工程师,主要从事天气预报技术研究。E-mail:nn172172@163.com。
