Multivariate prediction of PM10 concentration by LSTM neural networks
Air presence of particulate pollutants is an environmental problem with significant health issues. Monitoring their concentration is a key factor for the correct management of urban activities. In the smart cities scenario, the most fruitful tools for such application are sensor networks combined with machine learning techniques. In this work, neural networks are employed to forecast particulate concentration of air pollutants using a novel multivariate approach. We analyzed five years of data relating to PM10 concentration, studying the performance of different models based on the Long Short Term Memory paradigm, optimizing their hyperparameters accordingly. The tests show good results in terms of approximation and generalization capabilities, along with a sensible dependence on the weather conditions.