In order to solve the problem that some cities have serious PM_2.5 concentration data deficiency and can not get prediction model by training their own data, this paper proposes a method to use prediction model of similar cities to fill in historical data of target city. Based on the meteorological data, urban development data and PM_2.5 concentration data of 23 cities, an interpolation model of PM_2.5 daily concentration data is established based on self-organizing map (SOM) and gated recurrent unit (GRU) neural network and use it and raditional interpolation (linear interpolation and quadratic interpolation) methods are used to fill missing data respectively. The comparision of filling effects show that the SOM city matching model can accurately match the similar cities of the target city. For the missing data less than 5 days, the filling effect of traditional interpolation methods are better than that of GRU interpolation model. For the missing data more than 5 days, the GRU interpolation model is more competent to fill in the missing data for a long time.

为解决部分城市PM_2.5浓度数据缺值严重, 无法通过训练自身数据得到预报模型的问题, 提出用相似城市的预报模型实现目标城市历史数据的填补。依据23个城市的气象数据、城市发展数据和PM_2.5浓度数据, 建立基于自组织映射(SOM)和门控循环单元(GRU)神经网络的PM_2.5日均浓度数据插值模型, 并分别利用该插值模型和传统插值方法(线性插值和样条插值)对不同类型的缺值数据进行填补, 对比两者的填补效果。实验结果表明, 基于SOM神经网络的城市匹配模型可以准确地匹配出目标城市的相似城市; 当缺值数据少于5天时, 利用传统插值方法的填补效果优于GRU插值模型; 当缺值数据多于5天时, GRU插值模型更胜任长时间缺测数据的填补工作。