Long short-term memory artificial neural network approach to forecast meteorology and PM2.5 local variables in Bogotá, Colombia

In places with complex topography, the reproduction of atmospheric dynamics is challenging and resource demanding. Recently, machine learning has been successfully implemented to forecast pollution and weather variables. LSTM (long short-term memory) networks have the potential to improve the forecasting precision on different theoretical fields. Despite this advantage, they have not been widely used in the tropics for this purpose. This research aims to implement a LSTM to forecast PM2.5 and meteorological variables in a tropical mountainous city. The model was trained with 7 years of data from the local air quality monitoring network. The implemented model forecasts 42 days, evaluated using statistical scores and benchmarks. More than 95% of PM2.5 values, radiation (99%), air temperature (98%), relative humidity (95%), wind speed (94%), and the u-component (91%) have excellent or good benchmarks. The v-component and the wind direction got the lowest percentage of excellent or good values (50%). We compared our results with other models that have focused on forecasting these variables in similar places and observed that the LSTM approach results are the best, especially for PM2.5 and wind direction. We found its accuracy can be affected by rapid changes in the tendency of the data that do not occur as a consequence of the diurnal tide. The LSTM model was validated as a tool to predict meteorological variables and PM2.5 (24 h in advance) in a tropical mountainous city and can be used as a reliable input in air quality early alert systems.