Can Machine Learning Models be Used to Predict Pollutants based on Measured Other Pollutants?

In this paper, we investigate the use of machine learning (ML) models to estimate or predict concentrations of pollutants based on measured concentrations of other pollutants. Such models could be used in air quality index (AQI) detection systems to decrease the number of physical sensors in order to reduce overall and maintenance costs. Five different long-short term memory (LSTM) models were explored in the preliminary investigation. The most accurate model was then selected for further refinement via simple hyperparameter search. The final refined model was trained and tested on four different air quality datasets from four different countries. Simulation results indicate that prediction of pollutant concentrations based solely on measured concentrations of other pollutants is not accurate enough to warrant total sensor replacement with ML models. However, when the same ML models are provided as input past measurements of the predicted pollutant rather than previously predicted values, the prediction accuracy is excellent. We conclude that while ML models are not yet accurate enough to completely replace physical sensors, such models could be very helpful to provide predictions in situations of sensor failure and thus to guarantee continuous sensor fusion processes.