PM2.5 Concentration Prediction Based On Recurrent Neural Network Model

In recent years,our economy has achieved remarkable results.However,it has also paid a heavy price for resources and the environment.Most notably,air pollution is a critical problem to be urgently solved.With the worsening of air quality,a large area of hazy weather appears in our country.The suspended particles in the air,especially PM2.5,bring great inconvenience to people’s life and threaten people’s health.Therefore,if the concentration of PM2.5 can be effectively predicted,it will be beneficial to people’s travel.In this way,the necessary precautionary measures can be timely implemented.A majority of scholars use traditional statistical forecasting methods and machine learning methods for the numerical prediction of air pollutants.Most of these methods make some assumptions about the distribution of data.However,the forecasting results aren’t very accurate.In recent years,deep learning technology has been developed rapidly,of which the recurrent neural network can effectively predict the time series data.Therefore,in this paper,the LSTM(Long Short-Term Memory)recurrent neural network model was introduced into the field of air pollutant prediction.According to the hourly data of air pollution in Shenyang from 2013 to 2015,the concentration of PM2.5 in the next hour was predicted.The main contents of this paper included following two aspects:For the univariate PM2.5 own time-series data:based on the Keras interface,the BP neural network and the ordinary recurrent neural network were respectively established.A single hidden layer network structure was applied in both models,respectively studying the influence of hidden layer neurons’ number on network test error and fitting effect.And optimal network topologies were determined for both models.On the same test dataset,the average absolute errors of two models were 15.0566 and 13.7344 respectively.It can be found that the prediction results of ordinary cyclic neural network were better than those of BP neural network.For multivariable PM2.5 time series data:based on the historical data of PM2.5,this paper added the meteorological factors,and obtained the importance measure of meteorological factors through random forest regression.As a result,three meteorological factors were selected:humidity,air pressure and temperature.Coupled with the historical data of PM2.5,four characteristic variables were existed.Then,the LSTM model was established based on the Keras interface to predict the concentration of PM2.5 in next hour.The results showed that mean absolute error of LSTM model on the test set was 9.9495,predicting the PM2.5 concentration more accurately.