Meteorological influence on forecasting urban pollutants: Long-term predictability versus extreme events in a spatially heterogeneous urban ecosystem

With the escalation of heat- and pollution-related threats in cities across the globe, timely counteractions and emergency procedures are vital, which calls for accurate co-prediction of urban heat and air quality under both standard conditions and under extreme events. In this study, we used historical hourly data recorded at 9 sites across the Sydney metropolitan area to test the performance of long short-term memory (LSTM) forecasting architectures in predicting 5 urban pollutants based on different combinations of meteorological inputs and considering standard, bushfire, and pandemic lockdown conditions. We demonstrate that, in most cases and even in a fast-growing city, there is no significant benefit achieved by including extra predictors to temperature and humidity, when adequate forecasting techniques capable of learning long-term dependencies are used. Further, in agreement with previous studies, we provide evidence of ozone's higher responsiveness to all weather parameters and thus enhanced predictability and PM₁₀'s lower predictability as compared to all other considered urban pollutants. The prediction accuracy tends to be comparable between standard conditions and bushfire events. However, the predictability significantly declines under anomalies in anthropogenic patterns and urban metabolic rates as those recorded during the pandemic. The inclusion of local emission sources and anthropogenic factors in the input dataset is considered necessary for NO and PM₁₀ to properly predict urban air quality, especially under human-related extreme conditions.