Forecasting PM2.5 and NO2 Concentrations in Patras Using Low-Cost Sensors and Machine Learning

We present a machine learning methodology for forecasting next day’s PM2.5 and NO2 concentrations in Patras, Greece, using a sequence-to-sequence LSTM neural network architecture trained on data from a low-cost ENSENSIA sensor system and a meteorological station in Patras. The model integrates recent pollutant trends, temporal variables, and forecasted meteorology through an exogenous-aware encoderdecoder design with an attention mechanism. We used historical PM2.5 and NO2 concentration measurements from ENSENSIA and meteorological variables from a local weather station to build and train the model. Validation over December 2024 showed promising results, achieving a fractional error of 0.45 and bias of 0.01 for PM2.5, and 0.17 and 0.02, respectively, for NO2. The model underestimated PM2.5 peaks, highlighting limitations in modelling pollution events driven by atmospheric chemistry. Our results underscore the potential of combining low-cost sensing and machine learning for urban air quality monitoring.