Capability of an Elman Recurrent Neural Network for predicting the non-linear behavior of airborne pollutants

In this work an Elman Recurrent Neural Network (a type of Simple Recurrent Neural Network) is used for predicting the concentration of airborne pollutants O 3 , PM2.5 and PM10, which have a non-linear behavior, using data from Red Auto´ma´tica de Monitoreo Atmosfe´rico (RAMA), which is spreaded in the Zona Metropolitana del Valle de Me´xico (ZMVM). The study of PM10 and PM2.5 is important because, due to their tiny size, they can penetrate sensitive regions of respiratory system, among other important effects on human health, furthermore, it has been demonstrated that these have an important environmental impact. The study of ozone is important due its high toxicity. This pollutant has been responsible of several environmental contingences in Mexico City. An empirical method for imputing missing data using a series of linear regressions is proposed. A grid searching is used to find the best combination of some hyperparameters so that the variation of root-mean-square error between validation data and predicted data is minimized. A total of 144 experiments is developed measuring the validation root-mean-square error for each one, as well as root- mean-square error variation, in order to find the optimal combination. Two criteria are taken into account to evaluate the performance of network: root-mean-square error variation as mentioned before and evolution of metric values. This network showed a very good performance for ozone, with a maximum accuracy of 95.6 %, moderately good for PM2.5 with 46.4 and 28.6 % for PM10.