Sulfur dioxide concentrations forecasting using a deep learning model in Quintero, Chile

Abstract

Close to Quintero, a Chilean coastal city, located 160 km northwest of Santiago, a highly concentrated accumulation of industries generate high levels of atmospheric pollution which significantly affects the quality of life of its rural and urban population. The industrial complex, alongside other smaller industries, is home to an oil refinery, a copper foundry and 3 coal power plants. Sulfur dioxide (SO₂) frequently exceeds international and national standards in the area. Episodes of fainting and poisoning associated to high levels of SO₂ have been reported in Quintero. Due to this situation, it is highly relevant to develop a sulfur dioxide forecasting model which may be used as a tool to warn authorities and the local population about unfavorable air quality conditions. Three SO₂ forecasting models for the city of Quintero based on Machine Learning Techniques have been implemented: a Random Forest model, a Deep Learning Feed Forward model (DFFNN) and a Convolutional Long Short Term Memory (LSTM) Deep Learning model. The goal was to forecast the maximum of the hourly average value of SO₂ for the first 12 h of the following day based on information available during the present day. The LSTM model gives the best results with a 78 % accuracy.