EMD在广西季节降水预报中的应用

Application of EMD to Seasonal Precipitation Forecast in Guangxi

气候系统是一种耗散的、具有多个不稳定源的非线性、非平稳系统。该文利用支持向量机(SVM)算法在处理非线性问题中的优越性和经验模态分解(EMD)算法在处理非平稳信号中的优势,采用将EMD与SVM相结合的短期气候预测方法,并应用到广西季节降水预报中。选取广西88个气象观测站1957—2005年6—8月逐年降水量的距平百分率序列作为试验数据,通过EMD算法将标准化处理后的距平百分率序列分解成多个本征模态函数(IMF)分量和一个趋势分量,在分解中针对EMD算法存在的端点极值问题选择两种方法分别进行处理,对比得出极值延拓法效果更好。对每个分量构建不同的SVM模型进行预测,并通过重构形成最后的预测结果。试验中采用不经EMD处理的反向传播(BP)神经网络和SVM算法进行对比验证,结果表明:相对于直接预测方法,该文提出的方案均方误差最小,能够较为准确地反映出降水序列未来几年的变化趋势,具有更高的预测精度和较好的推广前景。

The climate system is a high order nonlinear system with dissipation.In recent years,the BP neural network algorithm and the Support Vector Machine（SVM） algorithm are applied widely in the short-range climate forecast for its superiority in handling nonlinear time series problem.Besides,the climatic time series are non-stationary,so the signal needs processing to improve its predication result.The Empirical Mode Decomposition（EMD） algorithm introduced by Huang is used to stabilize the climatic time series. Combined with the SVM algorithm,it＇s used for short-range climate forecast and applied to the seasonal precipitation forecast in Guangxi. The EMD algorithm decomposes non-stationary signal into several Intrinsic Mode Functions（IMF） components and a remainder with stationary.EMD algorithm doesn＇t provide a good solution for the endpoints extremes problem,and the extreme extending method is adopted as the endpoints continuation method for short-range climate forecast.Anomaly percentage of accumulated precipitation data are analyzed, which are observed at 88 meteorological observatories in Guangxi from June to August during 1957—2005.Using the EMD algorithm,the time series being standardized are decomposed into four IMF components and a remainder;then a SVM model is built for each component,and the forecasts are composed to the final forecast result.For comparison,BP neural network algorithm and SVM algorithm are adopted to forecast respectively without the EMD algorithm. Analysis on the predicted values and errors show that,without being processed with EMD,errors of the SVM algorithm are smaller than that of the BP neural network algorithm.So it proves that the generalization capability of BP is weaker than SVM when processing the small sample size problem,whereas SVM algorithm follows the structural risk minimization,and can coincidence the change trend better in condition of finite samples.It shows that the results of the EMD method combined with the SVM algorithm are more accurate.It illustrates that the EMD algorithm can reflect the regularity in different time scales of time series via decomposing into a collection of components with stationarity,which is more suitable for predicting with machine learning methods.The superiority of this scheme makes it widely applicable in precipitation forecast.