Air pollution is a global problem that directly affects the health of living beings; the World Health Organization (WHO) estimates that about 7 million of people die each year from exposure to polluted air. Having a prediction model for these air pollutants is an essential source of information for the proper prevention of health and life. There are many methods, and models for predicting air quality, almost all of them focused on large cities in the world. However, there are no models for cities considered underdeveloped and with high air pollution. Under this approach, the present project implemented an air quality prediction model for air pollutants (PM2.5, NO2, and 03). This is a proposal based on a method that combines a recurring neural network architecture LSTM and the increase of characteristics through a clustering process with K-means. The efficiency of our model was evaluated with the mean absolute error (MAE) and the mean square error (RMSE) and compared with machine learning algorithms: (Linnear Regression, K-Nearest, Random Forest, Decision Tree, and LSTM). Our proposed model (LSTM K-means) was more efficient than the traditional machine learning algorithms for regression; in the case of particulate matter (PM25) prediction, an MAE of 1.5 and RMSE of 2.39 was obtained, for Nitrogen Oxide (NO2) an MAE of 0.05 and RMSE of 0.06. For Ozone (O3), an MAE of 7.5 and RMSE of 9.81 was obtained, which are the minimum values compared to other algorithms.