Multi-Output LSTM-Based One-Step-Ahead Hourly Forecasting of PM₂.₅ and PM₁₀ at an Urban Station in Bogotá: A Two-Year Performance Analysis

Abstract

This study develops a one-step-ahead hourly forecasting model for PM2.5 and PM10 concentrations at an urban monitoring station in Bogotá, Colombia (Fontibón). The dataset, spanning from November 2023 to August 2025, comprises observations from a 22-month period retrieved from the Bogotá Air Quality Monitoring Network (RMCAB). Data pre-processing included physical range filtering and time-based interpolation to ensure a continuous time series. To prevent data leakage and ensure methodological rigor, the dataset was divided using a blocked chronological split into a training subset (10,501 observations) and a validation subset (2,500 observations), covering the period from April to August 2025. A multi-output Long Short-Term Memory (LSTM) network with 64 recurrent units and a 168-hour (7-day) sliding window was implemented to capture complex temporal dependencies. The model was optimized using Adam and Early Stopping, with weights restored from the best-performing epoch to prevent overfitting. Performance was evaluated in original physical units (µg/m³). For PM2.5, the model achieved a Mean Absolute Error (MAE) of 2.94 µg/m³, an RMSE of 3.81 µg/m³, and a coefficient of determination (R2) of 0.697. For PM10, the model attained an MAE of 11.87 µg/m³, an RMSE of 15.72 µg/m³, and an R2 of 0.702. Results indicate that the multi-output LSTM architecture effectively captures the non-linear dynamics of both pollutants simultaneously, explaining approximately 70% of the variance in the validation period. These findings establish a solid, reproducible proof of concept for integrating deep learning into urban early-warning systems, providing a scalable framework for air quality management in high-altitude Andean cities.