Prediction Methods Of Daily Temperature And Drought Index Using Support Vector Regression

Continuous hot weather which is so called heat wave is one of meteorological disasters caused by abnormal air temperature which also leads to the occurrence of drought. Along with global climate warming, the abnormal changes of temperature increase which threats directly to the healthy growth of economy and even endanger our life. Research on daily temperature prediction can predict the temperature changes to remind people adjusting life and production issues thus reducing or avoiding the loss of life and property. It also helps to improve the early warning mechanism of natural disasters. Drought prediction is one of hot topics in drought research field. Because the lack of mature theoretical system, drought index are used widely in drought forecast as one important tool to evaluate drought degree. Study on the prediction of drought index can improve the early warning capability of drought within the region which is an very effective way to change fighting against drought passively to forwardly.In this paper, modeling and predicting of daily temperature and drought index time series based on Support Vector Regression (SVR). The main works are as follows.(1) Study of daily temperature multivariate time series (MTS) prediction based on Local SVR. Constructing phase space of multivariate time-series with C-C method and minimum prediction error method computing delay time and embedding dimension to obtain samples. A way to extract nearest neighbors from each predictor’s sequences is used to build the local forecast model. Simulation based on daily temperature shows that the model has a better application value in short-term prediction compared with univariate time series and proposed extraction strategy can improve prediction accuracy effectively.(2) Applying MTS-LSVR method to predict Years-Max-Min Palmer Drought Severity Index (PDSI) time series which are extracted from monthly PDSI time series, and good performances are achieved.(3) Study of monthly drought index prediction based on SVR. Applying wavelet transform to decompose original drought index time series to get basis series and deviation series, series-direct-prediction method is presented to build SVR model for monthly drought index. Simulation shows series-direct-prediction model outperforms additive model. At the same time the applicability of Daubechies wavelets in prediction of Alabama North Valley PDSI time series is tested by experiments which prove four level decomposition with db5 has the highest accuracy.