DEEP RECURRENT NEURAL NETWORK FOR POLLUTION FORECASTING IN SMART CITIES

Abstract

Air pollution has been marked as one of the major problems of metropolitan areas around the
world. According to The Health Effects Institute (HEI), over 95% of the world's population is
breathing polluted air, which contributed to the death of 6.1 million people across the world in
2016. These health complications can be avoided or diminished through raising the awareness of
air quality conditions in urban areas, which could allow citizens to limit their daily activities in
the cases of elevated pollution episodes, by using models to forecast or estimate air quality in
regions lacking monitoring data, New automated paradigms base on artificial intelligent have to
be thought to improve the prediction performance. Many influencing factors make the prediction
complex. Traditional approaches depend on numerical methods to estimate the air pollutant
concentration and require lots of computing power. Moreover, these methods cannot draw
insights from the abundant data available, thus, Neural networks are recently applied in this
context due to their wide application. Deep recurrent neural network models provide a practical
approach to air pollution and air-pollution prediction. To address this issue, this paper puts
forward a deep learning approach using multivariate LSTM for the prediction of air pollution
concentration in smart cities. The City Pulse EU FP7 Project smart city data set was used in this
study. The proposed model was evaluated against state-of-the-art prediction techniques used in
pollution forecasting using RMSE, MAE, and R on Matlab 2018a. The proposed system model
for multiple pollutants forecast concurrently, which significantly increase the accuracy by 20%
and 36% in terms of RMSE and 26% and 47% in terms of MAE respectively.