The rapid increase in traffic, urbanization, and industrial expansion has all contributed to a decrease in air quality, which has a vital impact on both the long-term feasibility of the environment and the health of humans, particularly in industrialized nations. Numerous studies have explored using machine learning for air quality forecasting to reduce pollution. While shallow machine learning architectures offer less accurate forecasts, deep learning, a recent advancement in computational intelligence, has immense potential in predicting air quality. Deep learning frameworks can identify intricate correlations and patterns in data on air quality, resulting in more accurate and dependable predictions. Several aspects, including climatic conditions, emission sources, and geographical characteristics, may be considered by these models, which can help one better understand and anticipate air pollution levels. This research investigates deep learning applications' periodic changes in air quality. Hybrid deep learning methods utilize optimization, data decomposition, and correlation evaluation between PM2.5 particles and other factors to overcome limitations. This study contrasts various deep learning algorithms for forecasts of air quality and demonstrates that hybrid deep learning is more accurate compared to each model alone at predicting future periods of air quality. It proposes future research directions for the future generation of models. The literature summary provides valuable insights for academics seeking future studies in this field.