Radial-residual energy-aware deep learning framework for sustainable cloud–IoT intelligence in smart cities

Abstract

Smart city Internet of Things (IoT) networks is generating a continuous stream of heterogeneous sensor data that tends to require a timely analysis under the strict energy, latency, and computational constraints. Existing cloud-edge learning approaches have improved IoT intelligence, but they often treat feature learning, computation offloading, and task scheduling independently. Consequently, they have limited ability to adapt learning and processing decisions to the changing energy conditions of distributed IoT devices. This study proposes a Radial-Residual Energy-Aware Deep Learning (R2EADL) framework that jointly addresses these limitations through energy-aware feature learning and adaptive cloud-edge coordination. The proposed framework introduces radial attention to assign greater importance to informative and energy-sufficient sensor representations, while residual learning preserves feature information during deep representation learning. An Energy-Adaptive Cloud Coordinator dynamically determines whether computational tasks should remain at the edge or be transferred to the cloud according to device energy, data priority, and latency conditions. In addition, a reinforcement-learning scheduler adapts task execution to changing network workloads and resource availability. Experimental evaluation using smart city IoT datasets and heterogeneous network simulations has demonstrated that R2EADL achieves 87.9% energy efficiency, 131 ms average latency, 95.0% prediction accuracy, 93.1% task completion rate, and 9.3 tasks/s throughput, while the reinforcement policy converges within 102 epochs. Compared with the evaluated energy-aware and federated learning baselines, the framework has provided a more balanced mechanism for integrating feature intelligence, energy management, and cloud-edge task coordination.