Predict Model on Air Quality and Characteristic of Shenyang

In this paper, we use the data of air quality index (AQI) and other meteorological factors like temperature, humidity, dew point, wind direction, wind speed, pressure, and precip from 2013 to 2017 per hour to predict the future air quality in Shenyang. Now the Chinese government has not released the history of air quality in Shenyang so AQI data comes from the U.S. Department of State Air Quality Monitoring Program. Meteorological factors data come from Weather Underground. To begin with, we clean the data and convert the data into the format into a 3-dimension matrix. Then we first built a Fully Connected Neural Network (FCNN), after training, we find that the effect is not good, validation loss is much higher than training loss, over fitting happened. As the data is ranged by time so we generated the time sequence into the matrix and changed the Fully Connected Neural Network into Recurrent Neural Network (RNN). Specifically, we used Long Short-Term Memory (LSTM) layers to build the RNN. By using the Mean Absolute Error (MAE) loss function, we got about 27.5 for training loss, and similar to validation loss. Finally, we used Bidirectional Recurrent Neural Network (BiRNN) to make the prediction model with an MAE loss of about 8. Our result shows that according to the meteorological condition in Shenyang, in the future the AQI would still shaking from about 10 to 300. The variation mainly shows in different seasons. AQI in winter is much higher than it in summer, in spring and fall it shows stable, around 100.