广东开汛日期的多尺度物理统计预测模型

Multi-Scale Physical Statistic Prediction Model for Rain Season Onset Date in Guangdong Province

采用小波分析、Lanczos滤波器、相关分析、最优子集回归和交叉检验等方法,研究了广东开汛日期的多尺度变化特征及其与全球不同地区前期海温场、500hPa高度场的关系,建立了广东开汛日期的多尺度最优子集回归预测模型并进行了检验。结果表明,广东开汛日期存在显著的准6年和较明显的准17年周期振荡。广东开汛日期在年际和年代际变化尺度上与前冬海温场和500hPa高度场上共有20个显著相关区域,分别取对应时间尺度上显著相关区域的平均值作为预报因子,对相应时间尺度的广东开汛日期做最优子集回归,建立了相应的预测模型,以年际和年代际尺度上的预测值之和为广东开汛日期的预测值。所建立的预测模型具有较好的拟合效果,其中拟合值与实况值相差在5天以内的事件命中率为41.5%,10天以内的为60.4%。1951—2010年的交叉检验结果表明,广东开汛日期预测值和实况值之间的相关系数为0.33,通过了α=0.01的显著性水平检验。预测值与实况值相差在5天以内的事件命中率为26.7%,10天以内的为45.0%,因此,所建立的多尺度最优子集回归预测模型对广东开汛日期具有较好的预测能力。

In order to do better short-term climate prediction for the rain season onset date （RSOD） in Guangdong Province in each year, the multi-scale optimal subset regression prediction model for RSOD have already developed and tested through its multi scale variation characteristics and relationship with the antecedent SST and 500 hPa height in different time scales using the wavelet transform, Lanczos filter, correlation analysis, optimal subset regression and cross-validation. The results show that the RSOD in Guangdong Province has significant early trend, and exhibits interdecadal variation with 17.1-year period and interannual variation with 6.2-year period （after detrending）. There are twenty regions with signifi- cant correlation between RSOD and SST and 500 hPa height field in interannual and interdecadal time scales. The prediction models for interannual and interdecadal time scales are respectively constructed by optimal subset regression o~ RSOD in corresponding time scale with predictors defined by the mean of sig- nificant correlation region. The addition of interannual predictive value and interdecadal one is RSOD pre- diction. The model produces good regression effect. The percentage of hits of difference less than 5 days between regression and observation is 41.5%, and that less than 10 days is 60.4%. The result of cross validation analysis shows that the correlation coefficient between prediction and observation of RSOD is 0.33, which has passed the confident level of 99%. The percentage of hits of predictive error less than 5 days is 26.7%, and that less than 10 days is 45%. So the predictive model of multi scales optimal subset regression has good predictive ability for RSOD in Guangdong Provincc.