Air Quality Prediction Based On Neural Networks

Last decades,China has put its focus on economic development,which expedites industrialization progress,but,it brought severe environmental pollution.Among these pollutions,air pollution has drawn much attention of people,especially the haze.Although haze is deleterious to human's respirational system,it is hardly possible to uproot air pollution given the current technology of human,so the most efficient way to prevent people from damage of air pollution is prediction especially fine-grained prediction First,we try to predict air pollution based on ELM(Extreme Learning Machine),which has high training speed and good generalization performance.However,ELM cannot approximate complex function,more and more researchers began using deep learning models,such as CNN,RNN,to predict air pollution.Whereas,these models cannot treat input sequence and output sequence as sequence simultaneously,so we carried out research based on Seq2Seq(Sequence-to-Sequence)in order to exploit the sequential information.(1)To adapt ELM to fine-grained prediction,we proposed GBELM(Gradient Boosting Extreme Learning Machine).Its activation function was changed to ReLU(Rectified Linear Unit)and softplus and we also applied RBF(Radius Basis Function)kernel to it in order to improve its performance.Furthermore,to overcome its instability caused by determining the weight between input layer and hidden layer randomly,gradient boosting was used to integrate ELMs.The experiments turned out that ReLU and softplus could improve the performance of ELM obviously while RBF only improved little performance of ELM.With the help of gradient boosting,the performance of ELM was improved further.(2)Seq2Seq was used to predict air quality,but it had a very low training speed and the problem of error accumulation caused by recurrent prediction.To alleviate these two problems,we proposed n-step AAQP.In,n-step AAQP,a fully connected layer with position embedding replaced the RNN of encoder.In addition,n-step recurrent prediction was applied to decoder.Our experiments proved that n-step AAQP could achieve better result than original Seq2 Seq and accelerate the training process.The n-step recurrent prediction was useful to allay the error accumulation when the n was chosen appropriately.Finally,by comparing GBELM and n-step AAQP,we found n-step AAQP was more accurate on air quality prediction.Moreover,when facing sudden change in air quality,n-step AAQP had a better performance.