Lightweight Federated Learning with Genetic Optimization for PM 2.5 Forecasting in IoT Networks

This study presents a lightweight and privacy-preserving federated learning framework designed for resource-constrained IoT sensor networks, emphasizing efficient distributed computation across heterogeneous devices. The framework enables decentralized training of LSTM models directly on IoT nodes, eliminating the need for centralized data aggregation while ensuring full data privacy. To enhance computational efficiency and reduce communication overhead, a Genetic Algorithm-based model compression method is applied, pruning redundant weights and achieving approximately a 37% reduction in model size. The framework is evaluated on a real-world time-series forecasting task, achieving 66.3% classification accuracy across multiple categories, while also demonstrating low latency and high scalability in large-scale heterogeneous deployments. Furthermore, this approach supports real-time operation and effective management of hardware resource constraints, enabling practical deployment in distributed networks. These results highlight the potential of combining federated deep learning with evolutionary optimization to build efficient, secure, and scalable distributed IoT systems, providing a robust blueprint for future grid and edge computing applications.