基于时空优化LSTM深度学习网络的气温预测  

Air Temperature Prediction Based on the Optimized Spatiotemporal LSTM Deep Learning Network

作　　者：杨耘[1,2] 王彬泽 刘艳[3] 席江波 柏涵 王丽霞 吴田军 YANG Yun;WANG Binze;LIU Yan;XI Jiangbo;BAI Han;WANG Lixia;WU Tianjun(College of Geology Engineering and Surveying,Chang'an University,Xi'an 710054,China;Department of Cooperation of Chang'an University of State Key Laboratory of Geographic Information Engineering,Xi'an 710054,China;Urumqi Institute of Desert Meteorology of China Meteorological Bureau,Urumqi 830002,China)

机构地区：[1]长安大学地质工程与测绘学院,陕西西安710054 [2]地理信息工程国家重点实验室长安大学合作部,陕西西安710054 [3]中国气象局乌鲁木齐沙漠气象研究所,新疆乌鲁木齐830002

出　　处：《徐州工程学院学报（自然科学版）》2020年第2期44-49,共6页Journal of Xuzhou Institute of Technology(Natural Sciences Edition)

基　　金：长安大学中央高校基本科研项目(300102269205,300102269201,300102120201,300102269304);NSFC-新疆联合基金项目(U1703121)。

摘　　要：新疆天山山脉中段玛纳斯河流域及其周边地区气象观测站点稀疏且分布不均匀,导致对该地区积雪-融雪过程模拟所需的气温要素时空预测精度不高.针对这一问题,提出了基于时空优化长短时记忆(LSTM)深度网络的气温时空精准预测模型.首先,以该地区21个气象观测站点上2015年全年小时气象数据为数据源,利用Pearson相关性分析及多重共线性检验法选取了经度、纬度、风速、地表温度、相对湿度相关性较强的5个特征.其次,引入LSTM深度学习模型对气温等四个气象要素时间序列进行建模及预测,再引入后向传播(BP)神经网络对气象要素值进行优化并实现了将来逐小时气温的精准预测.最后,通过克里金插值(Kriging)制作了未来小时研究区气温空间分布图.对LSTM-BP模型预测精度进行分析,结果表明在研究区观测站点稀疏且分布不均匀情况下,利用提出的BP-LSTM模型预测的小时气温的均方根误差(RMSE)为2.37℃,比单独的LSTM模型降低2.21℃,比LSTM与多元线性回归组合模型降低0.3℃.LSTM-BP组合网络预测的绝对平均误差(MAE)也有所降低.对预测后的气温空间分布情况分析结果进一步验证了该模型的时空预测结果与实际情况一致.The meteorological observation stations in the Manas River Basin and its surrounding areas in the middle section of the Tianshan Mountains in Xinjiang are sparse and unevenly distributed,resulting in low accuracy of spatial-temporal prediction of temperature elements required for the simulation of snow-melt processes in this area.To solve this problem,a spatiotemporal accurate prediction model based on LSTM deep learning network is proposed.Firstly,taking the hourly meteorological data of 2015 from 21 meteorological observation stations in the region as the data source.Pearson correlation analysis and multicollinearity test method are used to select 5 features with strong correlation such as longitude,latitude,wind speed,surface temperature and relative humidity.Secondly,LSTM deep learning model is introduced to this model and predict the time series of four meteorological elements such as air temperature etc.,and back propagation(BP)neural network is introduced to optimize the value of meteorological elements and realize accurate prediction of hourly air temperature in the future.Finally,the spatial distribution map of air temperature in the study area in the coming hours is made by Kriging interpolation.The results from prediction accuracy analysis of LSTM-BP model showed that:the mean square error(RMSE)of air temperature prediction of the LSTM-BP network calculated is 2.37℃,which is 2.21℃lower than that of the LSTM alone,and 0.3℃lower than that of the combination of LSTM and multiple linear regression model.The absolute mean error(MAE)of LSTM-BP combined network prediction is also reduced.In addition,the analysis results of the predicted air temperature spatial distribution further verify that the spatiotemporal prediction results of this model are consistent with the actual situation.

关 键 词：长短时记忆(LSTM) 深度学习 气温 时空预测 玛纳斯河流域 

分 类 号：P457.3[天文地球—大气科学及气象学]