LSTM Networks for Particulate Matter Concentration Forecasting

Last years have witnessed an increasing interest in particulate matter present in the air and a greater awareness for its toxic effects. Fine particulate matter (PM2.5) concentration, i.e. particulate with diameter less than 2.5 µm, pose a severe risk to public health. This type of particulate matter due to its tiny dimensions can get deep into lungs and eventually even reach the bloodstream. Recent studies relate high levels of PM2.5 to lungs and heart disease. It can be very beneficial to be able to provide reliable forecasts for PM2.5 in the short-term horizon (24 h ahead). In this paper, a deep learning model based on a Long Short-Term Neural (LSTM) network has been evaluated. The dataset taken into consideration to validate experimentally the proposed model is publicly available at UC Irvine Machine Learning Repository and consists of hourly PM2.5 levels, from Jan 1st, 2010 to Dec 31st, 2015, for Beijing city, as well as atmospheric variables. The developed LSTM network is able to find useful patterns in previous values of PM2.5 that in conjunction to exogenous atmospheric variable provide reliable forecasting of PM2.5 in the short period 1–6 h, showing a progressive reduction in accuracy as the forecasting horizon increase. The accuracy of the proposed LSTM, expressed by the root mean square error (RMSE), has been evaluated using a rolling window validation methodology. Performance results, within the forecasting horizon of three hours, are interesting and qualify the proposed model as a “soft” indicative measurement method in the framework of current EU AQ Monitoring.