EBK和LSTM模型在气象干旱时空预测中的应用  

Temporal and Spatial Distribution Prediction of Meteorological Drought Based on EBK and LSTM Model

作　　者：张建海 张棋 许德合[2] 丁严 ZHANG Jianhai;ZHANG Qi;XU Dehe;DING Yan(Qinghai Hydrology and Water Resources Monitoring and Reporting Center,Xining 800001,China;North China University of Water Resources and Electric Power,Zhengzhou 450046,China)

机构地区：[1]青海省水文水资源测报中心,青海西宁800001 [2]华北水利水电大学,河南郑州450046

出　　处：《人民黄河》2020年第8期77-82,共6页Yellow River

基　　金：国家自然科学基金资助项目(51679089);2019年度河南省重点研发与推广专项(192102310257)。

摘　　要：为了检验气象干旱预测模型的有效性,利用1951—2017年河南省19个气象站逐月降水量数据,计算多尺度标准化降水指数(SPI),并建立了SPI序列长短时记忆神经网络模型(LSTM)。对模型参数进行率定和验证后,结合改进的经验贝叶斯克里金插值法(EBK),对河南省气象站多尺度SPI值进行时间序列预测和空间分布预测,并借助均方根误差(RMSE)、平均绝对误差(MAE)和决定系数(R2)对回归预测模型的有效性进行判定。结果表明:在时间序列预测方面,LSTM模型在1月、24月尺度SPI的RMSE值分别为1.116和0.392,MAE值分别为0.911和0.188,R2值分别为0.102和0.871,说明LSTM模型对SPI的预测精度与该指数的时间尺度有关,随着时间尺度的增大而逐渐提高;在空间分布预测方面,通过EBK得到的LSTM模型预测值在空间分布上与SPI观测值的空间分布十分相似,且能够在大尺度上较为精确地预测河南省旱情。In order to check the effectiveness of models for meteorological drought prediction,this paper applied the monthly rainfall data of 19 meteorological stations in Henan Province in the period of 1951-2017 to calculate the multiscale Standard Precipitation Index(SPI)and established the Long ShortTerm Memory Model(LSTM)of SPI sequences.After calibration and validation of the model parameters,it adopted the improved Empirical Bayesian Kriging(EBK)to predict the time series and spatial distribution of the multiscale SPI values of meteorological stations in Henan Province.The validity of the regression prediction model was determined by the Root Mean Square Error(RMSE),Mean Absolute Error(MAE)and Decision coefficient(R2).The results show that in terms of time series prediction,RMSE values of LSTM model in SPI are 1.116 and 0.392,MAE values are 0.911 and 0.188,R2 values are 0.102 and 0.871 respectively.The prediction accuracy of LSTM model in SPI is related to the time scale of the index and it increases with the increase of time scale.In terms of spatial distribution prediction,the predicted value of LSTM model obtained by EBK is very similar to the actual value of SPI in spatial distribution and can accurately predict the drought situation in Henan Province in a large range.

关 键 词：气象干旱 SPI LSTM EBK 河南省 

分 类 号：TV213.4[水利工程—水文学及水资源]