An Algorithm of Optimal Subset for Bayes Precipitation Probability Prediction Model

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 Naïve 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 Naïve 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 Naïve 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 Naïve 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-Naïve 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.