Influence of settings and predictors in neural network model performance: a Buenos Aires air quality case

Abstract

The design of Artificial Neural Networks is usually complex given the number of hyperparameters and predictors to be determined with crossed sensitivities. In the case of air quality, several antecedents seek to predict concentrations of pollutants, but generally, it is done with default Neural Network parameters and predictors selected with expert knowledge, which biases the results. In regions with scarce air quality measurements, this problem is even more complex. This study aims to explore and present a novel methodology for the design of a Multilayer Perceptron-type Neural Network for particulate matter prediction. Non-linear machine learning hybrid methods are implemented for the selection of predictors using a testing bench Perceptron and Self-Organizing Maps. The final model showed a better fit (correlation coefficient of 0.88 during the testing stage with new data and an root mean squared error of 1.8 µgm−3) than an expert model trained for the same study case and can be adapted for other regions and in other fields of study.