Generative adversarial network-based imputation and personalized federated learning with Convolutional Neural Network-Long Short-Term Memory for air quality index prediction

Air quality forecasting is a critical task for mitigating the adverse effects of pollution on public health and the environment. This research introduces an advanced Generative Adversarial Network based Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) Federated Framework (GC-LFF) that addresses the limitations of traditional centralized machine learning models, which struggle with data variability and privacy concerns. The proposed framework employs advanced convolutional neural networks and long short-term memory based architectures to effectively capture spatial and temporal dependencies in decentralized air quality data. Additionally, a Generative adversarial network-based imputation module is incorporated to manage missing or corrupted sensor data, ensuring completeness without necessitating centralized data sharing. Each participating city in GC-LFF trains its model locally using its air quality data and uses a weighted Federated Averaging technique to aggregate model updates on a central server. The final global model of each node is then customized to increase the accuracy of local forecasting. The GC-LFF framework was evaluated using individualized local data-driven fine-tuning across 26 metropolitan nodes. With a Mean square error (MSE) of 499.16, a Root Mean square error (RMSE) of 19.83, a Mean Absolute Error (MAE) of 14.41, and an R-squared ( R 2 ) of 0.72 after the removal of outliers. The model’s resilience and flexibility were validated using standard deviation analysis and confidence intervals. The results demonstrate its strong performance and scalability for practical implementation.