Deep Convolutional Self-Attention Network for Energy-Efficient Power Control in NOMA Networks

Abstract

In this letter, we propose an end-to-end multi-modal based convolutional self-attention network to perform power control in non-orthogonal multiple access (NOMA) networks. We formulate an energy efficiency (EE) maximization problem, and we design an iterative solution to handle this optimization problem. This solution can provide an offline benchmark but might not be suitable for online power control, therefore, we employ our proposed deep learning model. The proposed deep learning model consists of two main pipelines, one for the deep feature mapping where we stack our self-attention block on top of a ResNet to extract high quality features, and focus on specific regions in the data to extract the patterns of the influential factors (interference, quality of service (QoS), and the corresponding power allocation). The second pipeline is to extract the shallow modality features. Those features are combined and passed to a dense layer to perform the final power prediction. The proposed deep learning framework achieves near optimal performance, and outperforms traditional solutions and other strong deep learning models such as PowerNet and the conventional convolutional neural network (CNN).