Air quality forecasting in non-monitored urban areas through machine and deep-learning model

Air pollution poses a major environmental challenge, raising concerns about human health in urban environments. It leads to diseases such as asthma, exacerbates pulmonary conditions, and creates murky skies, lowering inhabitants’ quality of life. To quantify air pollution, cost-effective IoT (Internet of Things) devices are being deployed in cities, making air quality monitoring available for a wide range of end-users, including public administrations. However, full urban coverage is unfeasible, and awareness of the carbon footprint of IoT deployments is increasing. Therefore, new techniques are needed to maximize the value of IoT networks with reduced infrastructure. To address these challenges, this paper presents an air pollution analytical forecasting solution based on deep-learning/machine-learning techniques to estimate air quality in locations without deployed devices. Different combinations of well-known deep-learning models are compared with machine-learning techniques to determine the best approach for monitoring polluting gases and airborne particles based on well-defined evaluation metrics. Additionally, two new deep-learning techniques, Multipath-CNN-LSTM (M-CNN-LSTM) and Multipath-CNN-BiLSTM (M-CNN-BiLSTM), are proposed to conduct a more exhaustive comparison. Combinations of LSTM (Long Short-Term Memory) techniques give the best results, with different models working best for each pollutant. Specifically, LSTM was optimal for O 3 , and combinations of CNN (Convolutional Neural Networks) and BiLSTM (Bidirectional LSTM) worked best for NO 2 . GRU (Gated Recurrent Unit) was more efficient for PM 2.5 , and BiLSTM performed best for PM 10 . This demonstrates that the best strategy to accurately predict the time evolution of each pollutant’s behavior depends on the selection of the most suitable machine-learning or deep-learning technique. • M-CNN-LSTM and M-CNN-BiLSTM models are proposed for air pollution forecasting. • Forecasting is performed in locations without deployed sensor devices. • LSTM is optimal for O 3 , M-CNN-BiLSTM for NO 2 , GRU for PM 2 . 5 , and BiLSTM for PM 10 .