ARIMA-LSTM组合模型在基于SPI干旱预测中的应用--以青海省为例  

Application of a combined ARIMA-LSTM model based on SPI for the forecast of drought:A case study in Qinghai Province

作　　者：张建海 张棋 许德合[2] 丁严 ZHANG Jian-hai;ZHANG Qi;XU De-he;DING Yan(Qinghai Hydrology and Water Resources Survey Bureau,Joining 810000,Qinghai,China;North China University of Water Resources and Electric Power,Zhengzhou 450000,Henan,China)

机构地区：[1]青海水文水资源勘测局,青海西宁810000 [2]华北水利水电大学,河南郑州450000

出　　处：《干旱区地理》2020年第4期1004-1013,共10页Arid Land Geography

基　　金：国家自然基金(51679089、51609082、51709107)资助。

摘　　要：开展干旱预测是有效应对干旱风险的前提基础。利用1958-2017年青海省38个气象站点逐日降水量数据计算多尺度标准化降水指数(SPI),并建立了SPI序列自回归移动平均模型(ARIMA)、长短时记忆神经网络模型(LSTM)和基于二者优点提出的ARIMA-LSTM组合模型;对模型参数进行率定和验证后,利用所建立的模型,以西宁站点为例,对多尺度SPI值进行预测,借助均方根误差(RMSE)、平均绝对百分比误差(MAPE)和决定系数R2对所有预测模型的有效性进行判定。结果表明:ARIMA-LSTM组合模型在SPI1和SPI12的RMSE值分别为0.159 7和0.181 0,均低于ARIMA模型的1.265 4和0.293 3,说明ARIMA模型与ARIMA-LSTM组合模型对SPI的预测精度都与时间尺度有关,ARIMA模型的预测精度随着时间尺度的增加而逐渐提高;结合GIS并利用实测数据与模型的预测数据相比较说明ARIMA-LSTM组合模型相比于单一ARIMA模型的预测精度更高,且能够很好拟合不同时间尺度的SPI值。Drought prediction is a precondition to effectively mitigate the risk of drought.Daily precipitation data obtained from 38 meteorological stations in Qinghai Province,China in the period from 1958 to 2017 were used to calculate the multiscale standardized precipitation index(SPI).In addition,based on these data,the SPI sequence Autoregressive Moving Average model(ARIMA),Long Short-Term Memory model(LSTM),and ARIMA-LSTM combination model were constructed.After the calibration and verification of the model parameters,the model was used to predict multiscale SPI values using the Xining area as a case study.Moreover,the validity of all the prediction models was determined by root mean square error(RMSE)and mean absolute percentage error(MAPE).The results indicated that the RMSE values of the ARIMA-LSTM combined model in SPI1 and SPI12 were 0.1597 and 0.1810,respectively,which were lower than those(1.2654 and 0.2933)of the ARIMA model.This indicates that the prediction accuracy of the ARIMA and LSTM models for SPI was related to the timescale.Comparing the measured data(using GIS)to the data predicted by the models,the combined ARIMA-LSTM model exhibited a higher prediction accuracy compared to the single ARIMA model.In addition,the combined ARIMA-LSTM model showed an ability to fit the SPI values of different timescales.

关 键 词：干旱预测 SPI ARIMA-LSTM组合模型 青海省 

分 类 号：P426.616[天文地球—大气科学及气象学]