基于时间序列分解的降雨数据挖掘与预测

Rainfall Data Mining and Forecasting Based on Time Series Decomposition

研究建立了基于时间序列分解的神经网络模型,能对降雨时间序列挖掘并预测。(1)以桓台县1979-2018年的480组月降雨数据为例,将降雨时间序列分解为趋势项、周期项、突变项与随机项。(2)采用累积距平法、Mann-Kendall趋势分析法、Hurst指数法、特征点法方法进行趋势性分析;小波分析法进行周期性分析;Mann-Kendall突变检验法和Pettitt法进行突变性分析;采用自相关法和单位根法对随机项进行检验。(3)以1979-2014年的432组月降雨时间序列随机项为率定数据,2015-2016年数据为验证数据,分别建立NAR(Nonlinear Auto Regression)与NARX(Nonlinear Auto Regression with External Input)神经网络随机项预测模型,对2017-2018年月降雨数据进行预测,并与直接预测结果对比。结果表明:(1)桓台县1979-2018年月降雨量数据有微弱的上升趋势,预测未来将呈微弱下降趋势,其第一主周期是19(月),数据不存在明显的突变情况。(2)NAR神经网络所得2017-2018年的月降雨量预测值与实测值误差为16.79%。

A neural network model based on time series decomposition is established,which can mine and predict rainfall time series.(1)Taking 480 sets of monthly rainfall data of Huantai County from 1979 to 2018 as an example,the rainfall time series is decomposed into trend term,periodic term,mutation term and random term.(2)The cumulative anomaly method,Mann-Kendall trend analysis,Hurst index,characteristic point and other methods are used for trend analysis.The wavelet analysis method is used for periodic analysis.Mann-Kendall mutation test method and Pettitt method are used for mutation analysis;the random items are tested by autocorrelation method and unit root method.(3)Taking 432 sets of monthly rainfall time series random terms from 1979 to 2014 as calibration data,2015 to 2016 as testing data,the NAR and NARX neural network model is established to predict the monthly random items in 2017 to 2018 and compared with the direct prediction results.The results show that:(1)The rainfall data of Huantai County from 1979 to 2018 has a slight upward trend,and there will be a slightly decreasing in the future.The first main period is 19(month),without obvious mutation.(2)the error between the predicted and measured monthly rainfall in 2017 to 2018 by NAR neural network was 16.79%.