Air quality is a critical global issue, with air pollution posing serious environmental and public health risks, especially in urban areas where fine Particulate Matters (PMs) especially (PM2.5) is among the most harmful pollutants. Despite significant advancements in air quality monitoring and modeling, challenges such as data variability, computational demands, scalability, resolution constraints, and privacy concerns continue to limit the accuracy and effectiveness of current forecasting systems. This dissertation presents a novel approach to air pollution prediction in urban areas by integrating Artificial Intelligence (AI), spatiotemporal modeling, and privacy- preserving data collection techniques. The first major contribution of this research is the development of PMFORECAST, a temporal prediction model specifically designed to forecast PM2.5 levels. By utilizing advanced machine learning techniques and temporal attention mechanisms, PMFORECAST effectively captures temporal dependencies in pollutant concentrations, leading to highly accurate predictions for both short-term and long-term forecasting. Additionally, the model demonstrates significant multi-tasking capabilities. It achieves a notable prediction accuracy of 99.7% for 1-hour forecasts and 73.5% for 12-hour forecasts, representing substantial improvements over existing models in terms of precision and computational efficiency. Spatial and temporal data from underground sensor networks to predict PM2.5 concentrations across different geographic regions. The Graph Temporal LSTM (GT-LSTM) model employs Graph Convolutional Networks (GCNs) to capture the complex interactions between pollution sources and atmospheric conditions at ground level, while utilizing Long Short-Term Memorys (LSTMs), as described in the previous contribution, to model temporal dependencies. This approach provides a more refined understanding of pollutant dispersion over time and space. By operating with fixed zone resolutions corresponding to available data resources, the model ensures accurate and localized predictions. The third contribution is the design of a federated learning architecture called FEDAIRNET, aimed at enhancing air quality prediction using mobile sensor data while safeguarding user privacy. Traditional air quality monitoring stations are often constrained by limited spatial coverage and high costs, whereas mobile sensors offer a more flexible and granular data source. However, the collection of mobile sensor data introduces privacy concerns. FEDAIRNET addresses these challenges by distributing the learning process across multiple devices, ensuring that sensitive data remains on local devices while still contributing to global model updates. This decentralized approach not only improves prediction accuracy but also mitigates risks associated with centralized data collection, such as point-of-interest (PoI) attacks. The models presented in this thesis have been rigorously tested in real-world environments, demonstrating their potential to transform air pollution monitoring systems. The PMFORECAST model provides robust predictions of PM2.5 concentrations, making it valuable for public health interventions and environmental policies. The Spatiotemporal Model adds a critical layer of understanding by analyzing how pollutants behave across spatial and temporal dimensions, while the FEDAIRNET architecture ensures that privacy is protected as the use of mobile sensors becomes more prevalent. This research represents significant advancements in air pollution prediction by integrating AI-driven insights with privacy-preserving data collection techniques. Future work should focus on incorporating additional data sources, and refining hybrid models that combine temporal, spatial, and mobile sensing data. These innovations will contribute to more accurate, timely, and secure air pollution forecasting systems, ultimately helping to mitigate the harmful effects of air pollution on human health and the environment.