Using Deep Transformer Based Models to Predict Ozone Levels.

Ozone (O3) is an air pollutant that has harmful effects in human health when its concentration exceeds a certain level. Therefore, it is important to advance in methods that can appropriately predict O3 levels. In this paper we present a new model to estimate 4 h, 12 h, 24 h, 48 h and 72 h ahead O3 concentration levels. We rely on Deep Transformer Networks. Interestingly enough, these models were originally developed to be used in Natural Language Processing applications but we show that they can be successfully used in classification problems. In order to evaluate the usefulness of our model, we applied it to predict O3 levels in the centre of Madrid. We compare the results of our model with four baseline models: two LSTMs and two MLPs. Accuracy (Acc) and Balanced Accuracy (BAC) are the metrics employed to evaluate the goodness of all the models. The results clearly show that our Deep Transformer based Network obtains the best results.