融合Lamb-Jenkinson分型法和LSTM神经网络的PM2.5预测研究

PM2.5 Prediction Based on Lamb-Jenkinson Method and LSTM Neural Network

文章采用NCEP/NCAR逐日海平面气压场资料,利用Lamb-Jenkinson环流分型法对张家口大气环流进行分型,分析环流型与PM2.5质量浓度之间的关系,并针对PM2.5浓度的预测,提出一种融合Lamb-Jenkinson环流分型和LSTM神经网络混合模型的方法,即以环流指数为预测因子基于LSTM方法搭建PM2.5质量浓度的预测模型。结果表明:影响张家口地区的主要环流型有反气旋型、气旋型、偏北平直型、西南平直型、偏西平直型、东北平直型等、西北平直型、偏东平直型。PM2.5污染日出现的主要环流型为南气旋平直型、东南平直型、偏南平直型、偏东气旋型、西南气旋平直型、偏东平直型、气旋型等,而反气旋型和反气旋式平直环流型不利于污染出现。张家口地区的PM2.5污染与地面环流有着密切的联系,当存在PM2.5污染时,张家口地区处于日本海高压后部的均压场区域,污染越严重,日本海高压中心强度越强。模型预测结果的均方根误差为9.88、平均绝对误差为5.84、拟合优度达0.80,表明该模型具有一定的预报能力。

Based on the NCEP/NCAR daily sea level pressure data and the Lamb-Jenkinson atmospheric circulation type method is used to define atmospheric circulation types in Zhangjiakou,and the relationship between the circulation types and PM2.5 mass concentration was analyzed.This paper proposed a hybrid model combining Lamb-Jenkinson circulation type and LSTM neural network for PM2.5 quality forecasting,which is based on the LSTM method to construct the prediction model of PM2.5 prediction with the circulation index as the predictor.Results show that the main circulation types were A,C,N,SW,W,NE,NW and E type in Zhangjiakou.For PM2.5 pollution days,major types were CS,SE,S,CE,CSW,E and C types,while the anticyclone type and the anticyclone of straight circulation types are not conducive to pollution.PM2.5 pollution in Zhangjiakou area is closely related to the surface atmospheric circulation.When there is PM2.5 pollution,the area is at the back of the Japanese Sea highs.The more serious the pollution,the stronger the strength of the Japanese Sea highs.The root mean square error of the model prediction result is 9.88,the average absolute error is 5.84,and the goodness of fit is 0.80,which indicates that the model has certain forecasting ability.