基于深度学习的城市大气PM2.5浓度预测研究

Prediction of PM2.5 Concentration in Urban Atmosphere Based on Deep Learning

以气象因素作为PM2.5预测指标,对数据进行异常值处理和归一化处理,将时间序列数据集转变为监督学习数据集,建立多变量的基于注意力机制的GRU循环神经网络PM2.5预测模型预测PM2.5浓度,并与不含注意力机制的GRU循环神经网络预测结果对比。利用北京市2010年-2014年气象数据集进行仿真实验,实验证明两个模型都可以较好地预测PM2.5浓度值,可以为空气污染的治理工作提供数据支持。Attention+GRU循环神经网络对PM2.5浓度的预测比GRU循环神经网络更精确。

Taking meteorological factors as the prediction index of PM2.5,the data were processed with outliers and normalization,the time series da⁃ta set was transformed into the supervised learning data set,and the multivariable GRU cycle neural network PM2.5 prediction model based on attention mechanism was established to predict PM2.5 concentration,and the results were compared with those of GRU cycle neu⁃ral network without attention mechanism.Using the meteorological data set of Beijing from 2010 to 2014,the simulation experiment shows that the two models can predict the PM2.5 concentration value well,which can provide data support for air pollution control.The prediction of PM2.5 concentration by attention+GRU cycle neural network is more accurate than that by GRU cycle neural network.