基于VMD和LSTM对月气温和海温预测的对比研究  

Comparative Study on Monthly Air Temperature and Sea TemperaturePrediction Based on VMD and LSTM

作　　者：闫加海[1] 朱玉祥[2] 杨培芬[3] YAN Jiahai;ZHU Yuxiang;YANG Peifen(Shanxi Climate Center,Taiyuan 030006,China;Chinese Academy of Meteorological Sciences,Beijing 100081,China;Shanxi Meteorological Information Center,Taiyuan 030006,China)

机构地区：[1]山西省气候中心,山西太原030006 [2]中国气象科学研究院,北京100081 [3]山西省气象信息中心,山西太原030006

出　　处：《沙漠与绿洲气象》2024年第6期83-90,共8页Desert and Oasis Meteorology

基　　金：国家自然科学基金委员会—中国民用航空局民航联合研究基金(U2033207);中国气象局创新发展专项(CXFZ2023J027)。

摘　　要：在对比月平均气温和月平均海温距平两类时序数据长短期记忆网络(LSTM)开环、闭环模型预测效果基础上,引入变分模态分解(VMD)算法并开展组合建模,建立VMD-LSTM闭环模型、VMD-SLSTM开环模型、VMD-MLSTM开环模型。结果表明:VMD-LSTM闭环模型月平均海温距平预测效果较LSTM闭环模型提升明显,月平均气温预测效果提升有限;VMD-SLSTM开环模型、VMD-MLSTM开环模型预测效果较VMD-LSTM闭环模型和LSTM开环、闭环模型均有明显提升,两种开环模型平均RMSE较其他模型可减小1~2个量级,模型拟合数据与原数据相关性均在0.998 0以上。原因分析表明,VMD分解后各本征模态接近平稳信号,LSTM可有效捕捉其长程依赖性且预测时有新模态数据输入更新网络,其中VMD-MLSTM开环模型既有新模态数据更新网络又考虑了各模态间相互影响。Based on the comparison of the predictive performance of LSTM(Long Short-Term Memory Network)open-loop and closed-loop models for monthly mean air temperature and monthly mean sea surface temperature(SST)anomalies,this study incorporates the Variational Mode Decomposition(VMD)algorithm and conducts combined modeling,and establishes VMD-LSTM closed-loop model,VMD-SLSTM open-loop model,including VMD-MLSTM open-loop model.The prediction results show that the VMD-LSTM closed-loop model has a significant improvement in predicting monthly average SST anomalies compared to the LSTM closed-loop model,while its improvement in predicting monthly average air temperature is limited.The prediction performance of VMD-SLSTM open-loop model and VMD-MLSTM open-loop model is significantly improved compared to those of VMD-LSTM closed-loop model and LSTM open-loop and closed-loop models.The average RMSE of the two open-loop models can be reduced by one or two orders of magnitude compared to other models.The correlation between the model fitting data and the original data is above 0.9980.The analysis reveals that,after VMD decomposition,intrinsic models are close to stationary signals and LSTM can effectively capture its longrange dependence and update the network with new modal data input during prediction.The VMDMLSTM open-loop model not only updates the network with new modal data but also considers the mutual influence among modals.

关 键 词：VMD LSTM VMD-LSTM 气温 海温距平 时序预测 

分 类 号：P456.7[天文地球—大气科学及气象学]