Modeling and prediction of PM₂.₅ and PM₁₀ particles at urban stations in Bogotá using LSTM neural networks

Resumo

This article presents a proof-of-concept system for predicting hourly concentrations of PM₂.₅ and PM₁₀ in Bogotá using Long Short-Term Memory (LSTM) neural networks. The objective is to anticipate critical pollution episodes and support preventive decision-making in real time. Hourly data from the Bogotá Air Quality Monitoring Network were structured into 24-hour time windows and used as inputs to a 64-neuron LSTM architecture with two dense outputs for simultaneous estimation of both pollutants. The implementation in Keras/TensorFlow incorporated regularization techniques such as Early Stopping to improve model stability and reproducibility. Results show that while the model captures short-term fluctuations in PM₂.₅ with reasonable accuracy, performance for PM₁₀ remains limited, underscoring the exploratory nature of this study. The contribution lies in demonstrating the feasibility of recurrent neural networks for urban air quality forecasting and outlining pathways for future improvements, including the integration of meteorological covariates, larger datasets, and hybrid architectures such as CNN–LSTM and attention-based models. By positioning the work as a preliminary step, the study highlights opportunities to advance toward automated early warning tools aligned with current environmental regulations and the 2030 air quality reduction goals.