Monitoring pollution in megacities is challenging due to their vast urban expanses and numerous pollution sources. To address this, cutting-edge deep learning models are proposed for monitoring, managing, and forecasting pollution levels. In this study, data concerning six pollutants (O3, NO2, CO, SO2, PM2.5, PM10) and Air Quality Index (AQI) values were collected over a five-year period from three distinct zones within the Tehran megacity. The gathered data were incorporated into a dataset comprising 37,044 records, encompassing the concentrations of examined pollutants and the related AQI across various study points. The primary aim was to enhance the precision of deep learning models using a regional strategic separation approach. A frequency domain filter method was employed to address data noise. Four models—Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), Convolutional Neural Network with LSTM (CNN-LSTM), and Convolutional Neural Network with Bidirectional LSTM (CNN-Bi-LSTM)—were developed to predict AQI values. Evaluation showed disparities in pollution levels among the studied areas. The CNN-Bi-LSTM model exhibited the highest accuracy at 97 %, but challenges in predicting sudden fluctuations reduced accuracy to 95 % in some cases. However, the model consistently captured the underlying pattern of changes. Segregating megacity regions enhances accuracy in forecasting air pollution conditions.