<p>Air quality prediction is a critical tool for managing urban environments and safeguarding public health. Accurate forecasting of pollutants such as PM2.5 and PM10 is essential for timely warnings and effective pollution control. However, traditional prediction methods often struggle to capture the complex, nonlinear patterns in air quality data, leading to a high error rate. To address these limitations, this study introduces the optimal nonlinear regression self-organizing mapping recurrent network (BRSOMN), a hybrid deep learning model designed to improve air quality forecasting. The BRSOMN model integrates self-organizing maps (SOM) for dimensionality reduction and clustering, nonlinear auto regression with exogenous input (NARX) enhanced by crayfish optimization to capture temporal dependencies and external factors, and bidirectional long short-term memory (Bi-LSTM) networks to leverage historical data for future predictions. The model is trained and tested on a comprehensive dataset comprising PM2.5, PM10, wind speed, humidity, and gas concentrations (e.g., NO2, O3). Experimental results demonstrate that the BRSOMN model outperforms existing methods, achieving significant improvements in prediction accuracy. Performance metrics, including R² (0.89), RMSE (0.34), and NRMSE, indicate a substantial reduction in error rates compared to traditional and AI-based models. The proposed model also shows robust performance in capturing nonlinear trends and temporal dependencies, making it a valuable tool for real-time air quality forecasting. While the model achieves high accuracy, we emphasize that these results are context-specific and further validation is required to assess the generalizability across different regions and datasets. This study highlights the potential of the BRSOMN model for enhancing air quality management and protecting public health, particularly for vulnerable populations such as asthma patients.</p>