Air pollution forecasting using RNN with LSTM

With the advance of technology, it is increasingly exhaust emissions have caused air pollution. In particular, PM2.5 (Particulate Matter) has been proven that it has a great correlation with human health. Therefore, the detection and prediction of PM 2.5 air pollution is an important issue. There are countries around the world have built a variety of sensing devices for monitoring PM2.5 concentrations. There were also many studies have been constructed to predict and forecast various air pollution. Therefore, how to accurately forecast PM2.5 becomes an important issue in recent year. In this paper we propose an approach to forecast PM2.5 concentration using RNN (Recurrent Neural Network) with LSTM (Long Short-Term Memory). We exploit Keras, which is a high-level neural networks API written in Python and capable of running on top of Tensorflow, to build a neural network and run RNN with LSTM through Tensorflow. The training data used in the network is retrieved from the EPA (Environmental Protection Administration) of Taiwan from year 2012 to 2016 and is combined into 20-dimensions data; and the forecasting test data is the year 2017. We have conducted experiments to evaluate the forecasting value of PM2.5 concentration for next four hours at 66 stations around the Taiwan. The result shows that the proposed approach can effectively forecast the value of PM2.5.