The advent of Internet of Things (IoT) devices has led to an explosion in the amount of data generated at the network edge. Federated learning (FL) offers decentralized privacy-preserving model training; however, the high energy consumption of the client devices results in significantly higher cost and substantial sustainability issues. In this paper, we present an EA-FL framework for Edge-IoT devices that accommodates client selection, communication–computation trade-off and pruning/quantization techniques. We establish mathematical models to measure communication energy, computation cost, and overall performance. Experiments on Human Activity Recognition (HAR) and AMPds2 dataset shows that proposed approach provides up to 38% decrease in energy consumption with acceptable degradation model accuracy. The results demonstrate the applicability of EA-FL in terms of smart homes, industrial IoT and 5G environments. This framework follows the principle of green computing to federated learning on sustainable and intelligent systems.