基于CNN-Attention-BP的降水发生预测研究  

Application of CNN-Attention-BP to precipitation forecast

作　　者：吴香华[1] 华亚婕 官元红[1] 王巍巍[3] 刘端阳 WU Xianghua;HUA Yajie;GUAN Yuanhong;WANG Weiwei;LIU Duanyang(School of Mathematics and Statistics,Nanjing University of Information Science&Technology,Nanjing 210044;Nanjing Joint Institute for Atmospheric Sciences/Key Laboratory of Transportation Meteorology of CMA,Nanjing 210008;Experimental Teaching Center for Meteorology and Environment,Nanjing University of Information Science&Technology,Nanjing 210044)

机构地区：[1]南京信息工程大学数学与统计学院,南京210044 [2]南京气象科技创新研究院/中国气象局交通气象重点开放实验室,南京210008 [3]南京信息工程大学大气与实验环境教学中心,南京210044

出　　处：《南京信息工程大学学报（自然科学版）》2022年第2期148-155,共8页Journal of Nanjing University of Information Science & Technology（Natural Science Edition）

基　　金：国家重点研发计划(2018YFC1507905);国家自然科学基金(42075068,41975087);2020年江苏高校“大学素质教育与数字化课程建设”专项课题(2020JDKT032);南京信息工程大学2019年教改研究课题(共建共享的概率论与数理统计“金课”的探索与实践);南京信息工程大学数统学院本科专业建设项目。

摘　　要：在综合分析降水统计预测模型特点的基础上,提出一种基于Attention机制、卷积神经网络(CNN)和BP神经网络的CNN-Attention-BP组合模型,并对1961—2020年不同气候类型的长春站、白城站、延吉站夏季降水进行实证分析.首先,运用卷积神经网络对6—8月20—次日20时降水量、平均气压、平均风速、平均气温和平均相对湿度进行特征学习,利用Attention机制来确定气象影响因素对降水预测的权重;然后,使用BP神经网络进行降水发生预测,选用准确率、交叉熵损失函数和F1-score来综合评价CNN-Attention-BP组合模型的性能.最后,将单一的支持向量机、多层感知机和卷积神经网络模型与组合模型进行比较分析.结果表明,CNN-Attention-BP组合模型具有自主学习和关注更重要信息的特征,能够有效提高吉林省夏季降水发生模型的预测能力,在样本越均衡、降水频率越接近于0.5的站点,预测精度越高,准确率最高可达88.4%.CNN-Attention-BP组合模型的准确率相较于其他单一模型最高可以提高近17个百分点.Here, a CNN-Attention-BP model is proposed for precipitation forecast based on analysis of the characteristics of precipitation statistical prediction models, then empirical analyses are made on the summer rainfall in Changchun, Baicheng and Yanji stations of Jilin province for the period of 1961-2020.First, a Convolution Neural Network(CNN) is used to study the characteristics of precipitation, air pressure, wind speed, air temperature and relative humidity.Second, the attention mechanism is used to determine the weight of meteorological factors for precipitation forecast.Then a BP neural network is applied to predict the precipitation probability.And the performance of the proposed CNN-Attention-BP model is evaluated by accuracy, cross-entropy loss function and F1-score, which is compared with that of the support vector machine, multi-layer perceptron and convolution neural network model.The results show that the CNN-Attention-BP model is characterized by autonomous learning and paying more attention to significant information, as well as the improved forecasting performance in summer precipitation occurrence for Jilin province.Meanwhile, the proposed model would perform better with more balanced sample and precipitation frequency closer to 0.5,when accuracy would reach up to 88.4%.Compared with single models, the CNN-Attention-BP improves the forecast accuracy by 17 percentage points.

关 键 词：降水预测 卷积神经网络 Attention机制 BP神经网络 交叉熵损失函数 

分 类 号：P457.6[天文地球—大气科学及气象学]