ARIMA和LSTM方法长时间温度观测数据缺失值插补的比较  

Comparison of ARIMA and LSTM methods for interpolation of missing values of long-time temperature observations

作　　者：郑欣彤 边婷婷[3] 张德强[4] 贺伟 ZHENG Xintong;BIAN Tingting;ZHANG Deqiang;HE Wei(State Key Laboratory of Resource and Environmental Information System(Institute of Geographic Sciences and Natural Resources Research,Chinese Academy of Sciences),Beijing 100101,China;College of Resources and Environment,University of Chinese Academy of Sciences,Beijing 100049,China;College of Management,Beijing Union University,Beijing 100101,China;South China Botanical Garden,Chinese Academy of Science,Guangzhou Guangdong 510650,China)

机构地区：[1]资源与环境信息系统国家重点实验室(中国科学院地理科学与资源研究所),北京100101 [2]中国科学院大学资源与环境学院,北京100049 [3]北京联合大学管理学院,北京100101 [4]中国科学院华南植物园鼎湖山森林生态系统定位研究站,广州516065

出　　处：《计算机应用》2022年第S01期130-135,共6页journal of Computer Applications

基　　金：国家重点研发计划项目(2017YFD0300403)。

摘　　要：针对野外小气象观测站点半小时温度观测长时间数据缺失问题,结合较低频次的人工温度观测数据,采用时间序列分析和深度学习等方法,对缺失的半小时温度观测数据进行高精度插补。首先,选用深度学习数据插补中的序列-序列(Seq2Seq)方法,建立了适合高精度温度数据插补需求的编码-解码深度学习模型BiLSTM-I;然后,选用了传统的代表性方法,从时间序列回归分析——差分整合移动平均自回归模型(ARIMA)状态方程形式中,获取卡尔曼平滑状态估计方程的各项参数,由卡尔曼平滑估计实现对温度观测数据缺失值的插补。实验分析结果表明,所设计的BiLSTM-I深度学习气温插补方法要优于时间序列的双向递归插补方法(BRITS-I)。对缺失值时间窗口为30 d的测试集,测试结果中均方根误差(RMSE)为0.47℃,相较于BRITS-I得到的RMSE,精度提升了0.90;对缺失值时间窗口为60 d的测试集,RMSE为0.49℃,相较于BRITS-I得到的RMSE,精度提升了0.90;基于ARIMA状态模型的插补方法也有较高的精度,RMSE为0.75℃。最后,还分析了BiLSTM-I深度学习插补方法对不同温度缺失时间长度的适应能力,结果表明训练模型对不同的温度缺失时间长度具有泛化能力。Time series analysis and deep learning were used to interpolate the missing half-hourly temperature observations with high accuracy by combining the lower frequency of manual temperature observations,which addresses the problem of missing half-hourly temperature observations at meteorological observation stations in the field.First,The Sequence-to-Sequence(Seq2Seq)method which major in deep learning data interpolation was selected to establish the encoding-decoding deep learning model,named BiLSTM-I(Bi-directional Long Short-Term Memory),which is suitable for the demand of high precision temperature data interpolation.Then,the traditional representative method was done to obtain the parameters of the Kalman smoothing state estimation equation from the form of the state equation of time series regression analysis ARIMA(AutoRegressive Integrated Moving Average),and the interpolation of missing values of temperature observation data was realized by the Kalman smoothing estimation.The experimental analysis results show that the deep learning temperature interpolation method BiLSTM-I is better than BRITS-I(Bidirectional Recurrent Imputation for Time Series).For the test set with a missing value time window of 30 days,the result Root Mean Squared Error(RMSE)is 0.47℃and the accuracy is improved by 0.90 compared with BRITS-I;for the test set with a missing value time window of 60 days,the result RMSE is 0.49℃and the accuracy is improved by 0.90 compared with BRITS-I;the interpolation method based on ARIMA state model also has a high accuracy with RMSE of 0.75℃.Finally,the adaptability of BiLSTM-I deep learning imputation method to different temperature-absent time lengths was also analyzed,and the results show that the training model has the generalization ability to different temperature-absent time lengths.

关 键 词：气象观测数据 数据缺失 深度学习 时间序列分析 高精度插补 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]