Coverage Prediction for mmWave Mobile Communication Systems at the Edge: A Vision Transformer Approach

Coverage prediction is the pivotal quality-of-service indicator in cellular systems, which typically relies on propagation models for path loss estimation, leading to substantial computational overhead. In this paper, we present an efficient vision transformer (ViT)-based framework to directly capture correlations among terrain, interference, and coverage characteristics, bypassing the conventional path loss modeling. Experiments on urban scenarios demonstrate that ViT achieves the lowest prediction error against the ray tracing benchmark, outperforming COST 231 by more than 10% and exhibiting lower error than convolutional neural network (CNN)-based models, including AlexNet, ResNet, and a recent CNN for propagation prediction. Moreover, it also reduces computation time by up to two orders of magnitude compared to propagation models. This highlights the potential of our proposed method for low-latency coverage prediction with minimal prediction error in resource-constrained edge environments.