NINO3.4指数的多模式集合预报方法

基于贝叶斯模式平均方法(Bayesian Model Averaging)，发展了一个NINO3.4指数的多模式客观权重集合预报方法(简称OBJ)。该方法基于训练期内单个模式的预报结果，用线性回归订正单个预报的偏差，依据模式的预报效果估计单个模式的权重。利用2002年2月—2015年10月美国哥伦比亚大学国际气候与社会研究所(IRI)提供的7个单一模式对NINO3.4指数的预报结果进行OBJ试验，并采用均方根误差对多模式集合平均预报(简称ENS)和OBJ的预报结果进行检验和评估。结果表明，ENS的预报效果优于7个单一模式的预报效果，而OBJ预报效果优于ENS预报效果，其NINO3.4指数的均方根误差比ENS方法降低了4%。将单一模式预报结果按时间划分为训练期和预报期，利用独立样本估计OBJ的参数并进行预报试验，这些试验也表明，OBJ能进一步提高预报精度。 

A MULTIMODEL ENSEMBLE METHOD FOR NINO3.4 INDEX FORECAST

Based on Bayesian Model Averaging method, we developed an Objective Weighting Ensemble Method (OBJ) for NINO3.4, which is based on the prediction skill of individual models over a training period. In the OBJ method, the weights for individual models are estimated based on their past performances. Potential biases of individual models are corrected by linear regression of the individual model predictions with corresponding observations. In the conventional multimodel ensemble mean method (ENS), equal weights were used for individual models. Using predictions of NINO3.4 index of seven models from February 2002 to October 2015 for a total of 165 months, the weights of each model used in OBJ were estimated, then the skills of ENS and OBJ were compared using root mean square error. The results show that OBJ is better than ENS. The root mean square error of NINO3.4 index forecast decreases by 4% when OBJ is used, compared with that when ENS is used. Lastly, dividing the 165 months into two groups, the weights of OBJ were estimated using data in one group and hindcasts were verified using data in the other group. These results also confirm that the OBJ method is slightly more accurate than the ENS method.