Abstract: Tropospheric ozone impacts both human health and ecosystems. The prediction is crucial for public health awareness and implementation of air quality management strategies. This project aimed to create a forecasting model for ozone levels in Bogotá using data from the air quality monitoring network, including measurements of ozone precursors and meteorological variables. Machine learning models such as Convolutional Neural Networks and Birectional Long Short-Term Memory (LSTM) layers from the Tensor Flow Keras library, and the Python package Sklearn, were used to facilitate the categorization of artificial intelligence techniques. This results in a model that stands out for its ability to offer highly accurate forecasts by considering and quantifying the influence of precursors.The results showed a Spearman’s correlation coefficient larger than 0.6 (indicating a strong positive relationship) between the observed and predicted values in most cases. The Root Mean Square Error (RMSE) was below 10 μg/m3 (indicating a high level of accuracy) in all cases. Additionally, the Index of Agreement (IOA) was greater than 0.5 (indicating a good agreement) in all situations. This suggests that the model effectively and accurately reproduces the behavior and pattern of ozone. The proposed methodology is planned to be implemented operationally, serving as a complement to methods based on chemical transport models.

Keywords: Tropospheric ozone, Pollutant forecasting, Machine Learning, Convolutional Neural Networks, Birectional Long Short-Term Memory.