FORECASTING PM10 CONCENTRATIONS BASED ON MACHINE AND DEEP LEARNING

Abstract

A significant correlation exists between chronic exposure to a high level of particulate matter (PM) and an increase in health risks. To track and foresee particulate matter levels is necessary to reduce health risks and support the sustainable and healthy development of cities and communities. In this context, the aim of this study was chosen as implementing and comparing various Artificial Intelligence (AI) techniques in the prediction of particulate matter, and specifically PM10 concentration levels. The prediction approach implemented in the study was utilizing temporal forecasting models based on univariate time series of PM10 concentrations. Three different approaches were used for forecasting the PM10 concentrations. The first approach was based on Machine Learning techniques. The second approach utilized the Nonlinear Auto-Regressive Neural Networks (NARnets), and the third approach was focused on using Long-short term memory (LSTM) networks. The results of the study indicate that NARnets provide the most accurate results in the prediction of PM10 series with short term temporal dependencies.