Abstract

Predicting the urban air quality is necessary for maintaining public wellness, reducing the impact on the environment, and is helpful for sustainable urban planning. This research proposes a framework based on LSTM neural network for forecasting the air quality pollutants concentration, based on real-time pollution sensors data systematically combined with meteorological data. The methodology follows a defined data cleansing, natural science driven feature construction, and a multi-layer LSTM model designed to learn complex temporal patterns prevalent in the atmosphere’s components. The evaluation results indicate that the designed model significantly outperforms the comparatives ARIMA, Random Forest, and Support Vector Regression, reporting lower RMSE and higher R² for PM2.5, NO2, and O3 forecasts. The model’s performance is further validated by seasonal and trend analysis verifying representation of winter inversion persistent phenomenon, and the peak photolytic driving ozone synthesis, as well as stable performance simulation under pollution extremes for highly dynamic conditions. The results substantiate the LSTM model’s competitiveness for air quality monitoring and forecasting systems and for issuing alerts and public health advisory messages, and for policy directive on environment based on the collected data.