AIR QUALITY FORECAST SYSTEM USING NEURAL NETWORKS AND FORECASTED DATA FROM THE WEATHER RESEARCH AND FORECASTING MODEL
Across the globe, remarkable efforts are being invested to advance forecasts in various fields, including meteorology and air quality. Mexico City, a vast and densely populated urban area, often faces episodes of elevated pollution levels. To address these challenges, governmental authorities in the region require more precise air quality forecasts, among other critical information. This research presents a machine learning system tailored as an operational air quality prediction model for Mexico City. The proposed system leverages historical data from the Red Automática de Monitoreo Atmosférico (RAMA), an automated atmospheric monitoring network designed to persistently observe and document atmospheric conditions and a range of pollutants. In addition to the RAMA data, our model incorporates meteorological forecast data from a regional Weather Research and Forecast (WRF) model for Mexico City, enriched by contextual features such as the day of the week, current date, and time. The system is designed to integrate information from multiple stations, allowing it to compensate for missing data. The proposed system is a neural network trained to predict ozone levels for the subsequent 24-hour period in Mexico City by analyzing current and previous pollutant information from RAMA and seven meteorological forecast fields. This work evaluates the effects of bootstrapping the data, incorporating meteorological forecast information, and including current and previous pollutant time in the network’s performance. The model achieves a mean absolute error (MAE) of 9.5 ppb and a correlation index of 0.88 across all stations. For the top 10 stations, the MAE falls below 8.5 ppb, and the correlation coefficient exceeds 0.92.