Energy-Aware Federated Learning for AQI Pollutants Forecasting in Edge Networks

The criticality of the national air pollution control and frequent pollutant data acquisition failures demand forecasting of pollutants concentration. The monitoring edge stations face huge network traffic and communicates confidential air pollutants concentration. The Federated Learning (FL) by default offers privacy and reduces communication overhead. Also, the dependencies on centralized server can be avoided. This motivates us to build an efficient edge learning provisioned framework for forecasting the pollutants concentration. Since, the edge nodes are resource-constrained, these nodes need to be chosen dynamically for every round of training. To achieve this, we propose a FL using Optimal Node Scheduling (ONS) (Fed-ONS) based on network energy state and learning performance statistics. The Bidirectional Long-Short Term Memory (Bi-LSTM) is experimented and found to outperform other learning models considered. The error values are further reduced by 78.15% through optimal choice of node schedule with reference to random selection. Correspondingly, the network life time has been extended by 26.6%. The proposed ONS outperforms benchmark schemes such as Federated Averaging (FedAvg), Federated Averaging with proximal point optimization (FedProx) and Sign Stochastic Gradient Descent (signSGD). The mathematical analysis and experimental results emphasize that the proposed Fed-ONS mechanism enhances the training performance alongside with improved energy-efficiency.