Deep Learning Models for Low-cost Air Quality Sensor Calibration

Air quality monitoring systems have become an important part of urban areas due to recent attempts to monitor pollution levels to tackle problems such as climate change and population health risks. In recent years, research has been conducted of the utilisation of low-cost pollution concentration sensors to improve and expand on current air monitoring systems, as well as creating mobile systems that could be deployed in different scenarios. Although, the spread of Internet of Things (IoT) devices for monitoring systems brought the need of calibration between multiple different devices that could be found working inside the same network. This project explores the utilisation of Machine Learning and Deep Learning models to calibrate custom and Aeroqual sensors for \(PM_{2.5}\) and \(PM_{10}\) monitoring to an existing network from the city council in Cambridge, UK. For \(PM_{2.5}\), the collection with the custom sensor provided the highest accuracy when calibrated to the council one: Keras Regressor achieved an RMSE of 1.6240 and \(R^2\) of 0.8831, while with the data from Aeroqual a GRU Regressor achieved an RMSE of 1.9263 and \(R^2\) of 0.4867. On the other hand, collection with Aeroqual on \(PM_{10}\) concentration levels achieved an RMSE of 2.2087 and \(R^2\) of 0.6428 utilising RNN Regressor, while an MLP with Attention achieved a lower accuracy, with an RMSE of 4.9582 and \(R^2\) of 0.3297.