Prediction Of Air Quality Index Based On Recurrent Neural Network

Air pollution is becoming more and more serious today.It’s necessary to make an accurate prediction of air quality index(AQI).Because AQI and pollution concentration are time series,deep learning methods like Recurrent Neural Network(RNN)which can handle high dimensional data well are applied to AQI prediction.However,gradients explode or vanish problem is possible to appear when training RNN with backpropagation though time algorithm.Long Short Term Memory(LSTM)and Gated Recurrent Unit(GRU)neural networks can effectively solve this problem.Then,AQI can be predicted by these improved neural networks in this thesis.The structure of this thesis is organized as follows:(1)Prediction of AQI based on EMD-Stack-LSTM model.AQI data can be regarded as a set of signals owing to the nonstationary and nonlinearity.Firstly,running the classical EMD algorithm to decompose the original AQI data into several intrinsic mode functions and one residual signal.Then,running Stack-LSTM model to predict these IMFs and residual.Last,reconstructing the prediction value of these components so that we can get the prediction of AQI.This dissertation has proposed a EMD-Stack-LSTM model based on AQI prediction.Simulation experiment results has indicated that this algorithm is superior to the traditional recurrent neural network in prediction error and prediction accuracy.(2)Prediction of AQI based on EEMD-Bi-GRU model.For one thing,using Ensemble Empirical Mode Decomposition(EEMD)to improvement mode inclination.For another,running Bi-GRU algorithm to increasing running speed.Compared with other models,EEMD-Bi-GRU model is Confirmed that it can fit the nonlinear relationship between several air pollutants’ concentration and AQI well.(3)Prediction of AQI based on CEEMDAN-N-NCLSTM model.This dissertation has proposed CEEMDAN algorithm to decompose AQI to several smooth subsequences.it can improve noise interference problem of EEMD.With the abilities that nested network can access information of internal units selectively,that LSTM has high prediction accuracy,and that GRU has quick running speed,this thesis has proposed NNCULSTM model to predict subsequences.the advantages of high LSTM prediction accuracy and high GRU speed are preserved.To evaluate the performance of CEEMDAN-N-NCULSTM,we extract the air quality of Beijing which from January2014 to February 2020 and take it as the experimental object.Compared with several models,this model has superior prediction accuracy and lower evaluation indicators,it can be applied to actual air quality index prediction.