Abstract:

Based on the monthly mean temperature and precipitation at 160 weather stations in China during 1951-2011, 74 circulation indexes from National Climate Center and the monthly SST from NCEP/NCAR, combined with mean generating function (MGF), the prediction models of precipitation during the flood season in 5 stations of Liaoning Province were respectively established by using the partial least square regression (PLSR). And the effects predicted by PLS model on precipitation were tested. The result showed that the periodic factor of the prediction founded by MGF could weaken the correlation of the predictors in the statistic model to some extent. The root mean square error (RMSE) of mean precipitation during the flood season in Liaoning Province predicted by PLSR model considering the periodicity from 2002 to 2011 was reduced by 10.0 mm. The effect predicted by PLSR model on precipitation during the flood season was much more efficient in comparison with that predicted by step wise regression (SWR) model owing to better resolve about the multi-correlation problem. The mean prediction skill score of precipitation during the flood season in 5 stations from 2002 to 2011 was 72.6%, which improved by 10.3% than that of SWR model.

Key words: partial least square regression, mean generating function, precipitation prediction during the flood season

摘要：

利用1951～2011年中国160个气象站逐月降水、温度、74项环流指数和NCEP再分析海表温度资料，采用偏最小二乘回归（PLSR）方法，结合均生函数构造预报量周期性因子，建立辽宁省汛期平均降水量及其5站（沈阳、朝阳、营口、丹东和大连）汛期降水量预测模型，并进行预测效果检验分析。结果表明：采用均生函数构造预报量周期性因子，在一定程度上弥补了气候预测统计模型高相关性因子的不足，从而使辽宁汛期平均降水量PLSR模型的试报均方根误差降低约10 mm。PLSR模型由于较好地解决了预报因子之间的多重相关性问题，其预测效果较逐步回归模型有明显提高，对2002～2011年辽宁5站汛期降水量试报的Ps评分平均值为72.6%，比逐步回归模型提高了10.3％。

关键词: 偏最小二乘回归, 均生函数, 汛期降水预测