计算机科学
变压器
稳健性(进化)
实时计算
度量(数据仓库)
电子工程
算法
人工智能
控制理论(社会学)
工程类
出处
期刊:IEEE International Symposium on Dynamic Spectrum Access Networks
[Institute of Electrical and Electronics Engineers]
日期:2026-05-11
卷期号:: 225-232
标识
DOI:10.1109/dyspan69846.2026.11571154
摘要
Accurate pathloss (PL) prediction is fundamental to the planning and optimization of modern wireless systems. Encoder–decoder convolutional neural network (CNN) architectures have been widely adopted for data-driven radio map generation; however, their inherently local inductive bias can under-represent long-range, path-dependent effects governed by the transmitter–receiver (Tx–Rx) geometry, which may result in large outliers. In this work, we propose a physics-informed transformer model for data-driven outdoor PL prediction. The transformer casts PL prediction as sequence learning along the Tx–Rx terrain profile using structured tokens and a shared model for line-of-sight (LOS) and non-line-of-sight (NLOS) regimes. As a case study, we evaluate the irregular-terrain PL prediction performance using a large-scale dataset generated by the Terrain-Integrated Rough-Earth Model (TIREM). Evaluated under propagation parameter set identifier (PID) variation (fixed location) and location identifier (OID) variation (fixed parameters), the transformer solution consistently reduces the tail of the error distribution across most cases, improving robustness relative to a strong CNN-based UNet (CNN–UNet) baseline that predicts PL maps from terrain and parameter channels.
科研通智能强力驱动
Strongly Powered by AbleSci AI