Improving raw readings from low-cost ozone sensors using artificial intelligence for air quality monitoring

Abstract. Ground-level ozone (O3) is a highly oxidizing gas with very reactive properties. It is harmful at high levels and is generated by complex photochemical reactions when primary pollutants from the combustion of fossil materials react with sunlight. Thus, its concentration indicates the activity of other air pollutants and plays a crucial role in smart cities. With the growing interest in high-resolution air quality (AQ) monitoring, low-cost ozone sensors present an interesting alternative, although they lack accuracy and suffer from cross-sensitivity issues. In this context, artificial intelligence techniques, particularly ensemble machine learning (ML) models, can improve the raw readings from these sensors by incorporating additional environmental information to minimize inaccuracies and nonlinearities, as well as by including metadata to account for sensor aging effects and improve the models based on road traffic patterns. In this paper, based on the low-cost ZPHS01B multisensor module with nine sensors, we analyze, propose, and compare different techniques using four ML models in a low O3 concentration scenario (mean value of 55.72 µg m−3). We carried out a thorough exploratory data analysis process to extract the main features (variables) and performed hyperparameter optimization for the different models. As a result, we reduced the estimation error by approximately 94.05 %. In particular, using the gradient boosting algorithm, we achieved a mean absolute error (MAE) of 4.022 µg m−3 and a mean relative error (MRE) of 7.21 %, outperforming related work while using a module approximately 10 times less expensive. To carry out this work, we generated two datasets in the city of Valencia (Spain), at two different locations with the same characteristics (close to the ring road but separated by 4.1 km), of 165 and 239 d.