局部均值伪最近邻算法在降水预报中的应用

分析北京地区日降雨量资料发现,相较于其他降雨事件,大雨或暴雨事件发生的次数较少,因此该地区的降水量预报属于样本不均衡问题。在样本不均衡的情况下,K最近邻算法的分类误差率将会大大提高,这也就使传统的基于K最近邻算法的降水量预报方法的应用受到了限制。针对北京地区降水量预报这一样本不均衡问题,应用局部均值伪最近邻算法构建了北京市的降水量预报模型。该方法利用北京地区日降雨量资料和美国国家环境预报中心全球格点资料,将降雨量作为类,将美国国家环境预报中心全球格点资料的各种因子场作为天气样本特征,计算得到不同天气样本在所有类中的局部均值伪最近邻,通过决策规则实现最优分类。利用提出的降水预报模型对北京市2010年6—8月进行了24 h降水预报,实验结果表明,提出的预报方法对于降水等级预报的预报准确率以及晴雨预报的TS评分、正样本概括率、空报率和漏报率均优于传统的K最近邻预报方法,该方法具有较好的预报效果。

Application of local mean-based pseudo nearest neighbor algorithm in precipitation forecast

After analyzing daily precipitation data in Beijing City, the number of heavy rainfall events or torrential rainfall events, as compared to the number of other events such as no rain events, light rain events and moderate rain events, is significantly few. Therefore, the precipitation data of precipitation forecast in Beijing city is imbalance. In the case of the precipitation data with an uneven distribution, the forecast performance of the traditional k-nearest neighbor algorithm based precipitation forecast method will degrade dramatically due to the classification error rate of the k-nearest neighbor algorithm rising greatly. To overcome the problem which is the sample size of the precipitation forecast in Beijing city is not balanced, the precipitation forecast model is established using Local Mean-based Pseudo Nearest Neighbor (LMPNN) algorithm. Using the observed precipitation data in Beijing as the class and factors of the National Centers for Environmental Prediction-National Center for Atmospheric Research reanalysis dataset as the feature of the sample, we firstly calculated the local mean-based pseudo nearest neighbor of the sample in each class and then obtained the class label of the sample by the LMPNN rule. The extensive experiments of 24 hour precipitation forecast of Beijing city from June 1st to August 31th in 2010 are conducted. The proposed approach could get more accuracy, threat score, summary alarm rate, false alarm rate and missing alarm rate than that of the precipitation forecast method based on k-nearest neighbor algorithm. The comprehensively experimental results suggest that the proposed approach is effective.