基于时间尺度分离的斗门开汛日期预测模型

A Prediction Model for Rain Season Onset Dates in Doumen District Based on A Multiscale Decomposition

采用小波分析、低通滤波、相关分析、逐步回归以及最优子集回归等方法建立斗门开汛日期的预测模型。将开汛日期序列分解为年际变化和年代际变化2个分量,分别研究其与前期冬季500h Pa高度场、850 h Pa风场、海平面气压场和海温场的相关性,并将显著相关区域作为初选因子。通过逐步回归将因子数限制在10个以内,最后应用最优子集回归建立预测模型。结果表明:这种分离时间尺度的回归预测模型具有较好的拟合效果,计算得出的回归序列与原距平序列相关系数为0.81,两者同号率为77.8%,其中相差±7 d以内年份约占42.2%;相差±15 d以内的年份约占75.6%。最后对2012—2014年开汛日期进行试报,其中2012和2014年试报效果较好,误差分别为-5.89和-10.45 d。

Using the wavelet analysis, lowpass filtering, stepwise regression and optimal subset regression, we constructed a prediction model for rain season onset dates in Doumen. The time series of onset dates was decomposed into interdecadal and interannual scales, and their relationship with the 500 hPa height, 850 hPa wind, sea level pressure and sea surface temperature in precedent winters was analyzed respectively to select primary predictors. Then stepwise regression is applied to limit the number of predictors to be within ten and finally a prediction model was constructed by using optimal subset regression. The results show that the prediction model has a good regression effect. The correlation coefficient between the regression and observation is 0. 81, the percentage of hits of the same-sign symbol is 77.8%, the difference of less than 7 days is 42. 2% and that of less than 15 days is 75.6%. The forecast results of 2012 and 2014 are good, and the prediction errors are - 5.89 and - 10. 45 days respectively.