一种Bayes降水概率预报的最优子集算法

An Algorithm of Optimal Subset for Bayes Precipitation Probability Prediction Model

MOS预报最优子集模型,通过消除数值模式系统性误差,可最大程度地提高其预报技巧。为了建立Nave Bayes降水最优模型,利用2008 2011年T511数值预报产品和单站观测资料,对介休、运城、丰宁3个站Nave Bayes降水概率分级预报模型进行研究。通过设计恰当的适应度函数,提出了一种用遗传算法搜寻Nave Bayes模型最优子集的计算方案,得到了3个站的最优子集模型。结果表明:最优子集的拟合效果明显高于普通初始子集,能够显著提升数值模式在单站的预报技巧。最优子集模型主要通过降低数值模式空报率提高单站晴雨、小雨预报效果,通过小幅提高正确次数和降低空报次数改善对中雨预报效果。

Based on numerical prediction products, a model output statistic （MOS） for precipitation forecast of an observatory is set up which contains the model output rainfall as one of predictors. The model can remove the systemic error of numerical prediction on precipitation, so it improves the precipitation prediction skill to certain degree. But for a given amount of predictors, a problem to solve is how to select the optimal subset to improve the prediction skill especially in operational weather forecast. In order to construct a Naive Bayes precipitation probability prediction model on the precondition of the best performance from optimal subsets, using T511 model products and their 13-hour to 24-hour forecast corresponding observation of precipitation from 2008 to 2010 at three observatories, namely Jiexiu, Yuncheng and Fengning, the classificatory Naive Bayes models on precipitation probability are developed and valuated. Different from the treatment of classic optimal subsets regression which enumerates the optimal subset one by one under the rule of couple score criterion （CSC）, a Naive Bayes model using genetic algorithm to search the optimal subset from a great many of subsets is presented. Model follows artificial intelligence searching characteristics. The genetic algorithm is established through the construction of gene bit-series from binary encoding method, and the introduction of a fitness function with cause. Considering the elimination of non-existing affair samples for the weather of low probability, two models are built based on genetic algorithm and Naive Bayes model. The essential difference between two kinds of models is the fitness functions they use; One uses the accuracy of precipitation as fitness function, and it is called genetic algorithm-Naive Bayes forecasting model type 1, GA-NB1 in brief; the other one uses threat score as fitness function, and is called GA-NB2 accordingly. The models are evaluated by prediction tests with dataset ranging from July to September in 2011. Results indicate that simulated results of optimal subset are much superior to those of ordinary initial subsets. Both GA-NB1 and GA-NB2 can improve T511 model precipitation accuracy by 19% on precipitation occurrence, threat scores are improved by 0.16 and 0.13 on drizzle and moderate precipitation, respectively. The prediction for precipitation occurrence and drizzle is enhanced by the optimal subset model because they effectively reduce the false alarm rate of numerical model, by more than 19 times during the period. The cause for improving moderate rain prediction includes two aspects. A slight increase in the amount of correct forecast and decrease of false alarms.