多基站协同训练神经网络的PM2.5预测模型

PM2.5 prediction model based on multi-station co-training neural network

针对通过数值方法对PM2.5进行预测已经取得了良好的效果,但相关模型重视时间影响因子而对空间影响因素的关联性考虑不足的问题,该文提出了多基站协同训练长短时记忆网络预测模型。该模型以时空数据作为输入,并将多个基站数据进行协同训练。MC-LSTM网络通过采用多基站共享参数的方式,减少了需要训练的网络复杂度,减轻了网络过拟合的风险。利用MC-LSTM网络对北京市21个监测基站数据进行了处理,结果表明:MC-LSTM网络能够同时对各个基站的PM2.5浓度进行预测。

The prediction of PM2.5 concentration is one of the hot topics in the field of Environmental Science.In recent years,numerical methods have been used to predict PM2.5,but the related models focus on time impact factors while lack of consideration of spatial factors.Aiming at the problems of previous research,this paper proposes MC-LSTM（Multi-station Co-training LSTM）network prediction model.The model takes spatio-temporal data as input and trains each station＇s data together.In addition,MCLSTM network reduces complexity of network by sharing parameters,thus reducing the risk of overfitting.Finally,the effectiveness of the model is demonstrated by using 21 monitoring stations in Beijing.The MC-LSTM network can predict the PM2.5 concentration of each station at the same time,which improving the model＇s practicability.