Currently, air pollution is a topic of high importance in society due to its harmful effects on human health and the environment. Among the various air pollutants, PM2.5 (particulate material with diameter less than 2.5 micrometers) is relevant because high concentrations in the air can trigger respiratory, vascular or even lung cancer problems to people that live in contamined areas. Currently, the prediction of concentration of this material in Santiago de Chile is typically based on statistical methods or classic neural networks. In this work, we propose a model for the prediction of PM2.5 concentration and its critical events in Santiago de Chile through the use of recurring LSTM and GRU networks. In particular, data from the air quality monitoring stations located in different parts of the city of Santiago is used to predict the level of pollution by hours. The work describes the experiments carried out, with particular emphasis to the preprocessing of the data for its importance in the identification of the model. The obtained experimental results show that the performance of the GRU network slightly exceeds the LSTM network, reaching a coefficient of determination greater than 0.78 in independent set of testing data, while the threshold prediction in both networks exceeds 0.87 of R2 in the same testing set. As future work, we intend extend the proposed models to a spatial prediction throughout the city of Santiago.