基于深度学习的气象要素预测方法  

METEOROLOGICAL ELEMENTS FORECASTING METHOD BASED ON DEEP LEARNING

作　　者：马景奕[1] 刘维成 闫文君 MA Jing-yi;LIU Wei-cheng;YAN Wen-jun(China Meteorological Administration Meteorological Cadre Training Institute Gansu Branch,Lanzhou 730020,China;Lanzhou Central Meteorological Station Lanzhou 730020,China)

机构地区：[1]中国气象局气象干部培训学院甘肃分院,甘肃兰州730020 [2]兰州市中心气象台,甘肃兰州730020

出　　处：《热带气象学报》2021年第2期186-193,共8页Journal of Tropical Meteorology

基　　金：国家自然科学基金项目(41505036)资助。

摘　　要：针对气象预测内容繁多且影响因素多样的问题,提出了一种基于长短时记忆(LSTM)的气象预测方法。方法能够对繁杂的气象数据进行自动预处理,提取相应的特征信息。通过神经网络的前向训练、长短时记忆反馈学习,经过多隐藏层地自主训练,对能见度、温度、露点、风速、风向以及压力气象信息实现准确预测。通过实验以及与经典机器学习预测方法的比较,验证了本文方法在气象预测中的有效性,进一步提升了气象预测的准确性,各项预测值的均方检验误差平均值为0.35。Aiming at the problems of various weather forecasting content and various influencing factors,a weather forecasting method based on long short-term memory(LSTM)is proposed.The method can automatically preprocess the multifari-ous meteorological data and extract the corresponding characteristic information.Through the neural network forward training,long and short time memory feedback learning,through the multi-layer neural network autonomous training,to achieve the visibility,temperature,dew point,wind speed,wind direction and pressure meteorological information accu-rate prediction.Through simulation experiments and comparison with classical machine learning prediction methods,we verify the effectiveness of the method proposed in this paper in weather prediction,and further improve the accuracy of weather prediction.The mean square test error of the predicted values is 0.35.

关 键 词：气象预测 深度学习 神经网络 长短时记忆 

分 类 号：P456[天文地球—大气科学及气象学]