Real-Time Air Pollution Monitoring Using Machine

Abstract

According to World Health Organization (WHO), the air in Lahore, Pakistan, has an annual average of 68 µg/m3 of PM 2.5 particles which are 6.8 times more than safe levels recommended by WHO. In December 2016, the amount of PM2.5 went above 100 µg/m3 which was more than 10 times safe levels set by WHO. Currently, there is no network for real time air pollution prediction available in Lahore, nor Pakistan. In this research, an Internet of Things (IoT) based low cost/low power outdoor air quality monitoring system has been developed. We developed the wireless sensor nodes for the monitoring and prediction of the PM2.5 particles in the air. The nodes comprised of different sensors for measuring PM2.5, SO2, O3, CO, NH3, NO2, humidity, and temperature values and Long Range (LoRa) transceiver for uploading the data on the cloud. The coverage range of LoRa gateway has been tested in Lahore, Pakistan and it is found that coverage range of the sensor node is around 2.7 KM in densely populated areas. Five different artificial neural network models are used for the prediction ahead of 1hr, 2hr, 3hr, 4hr, 5hr and 6hr. We obtained the accuracy of 99% for tested model of LSTM. The accuracy of LSTM model for 1hr ahead is 13% more than FFNN, 2% than Elman NN, 0.41% than NARX and 0.45% than layer recurrent. The accuracy of LSTM model to predict 6hr ahead is above 95%. For 6hr ahead prediction, accuracy of LSTM is 15% better than FFN, 12% than Elam NN, 7% than layer recurrent and 9% than NARX NN. Results show that the LSTM model outclass all other models. The computational time of LSTM model to predict 6hr ahead is 150.22565msec. While FFN takes 30 msec, Elman NN takes 60msec, NARX takes 80msec and layer recurrent take 90 msec. The 1hr ahead predicted model of feed-forward NN embedded in all nodes and placed at different places for real-time testing. The dataset collected at the server side analyzed and it is found that our air pollution monitoring system is capable of accurately predicting the future (1 hour ahead) concentration of PM 2.5 in outdoor air. Its accuracy for 1hr ahead prediction is up-to 90% and its algorithm takes 0.642 micro seconds computational time for the prediction level of PM2.5.