Performance Comparison Study of Neural Network Architectures in Radio Frequency Map Prediction for Wireless Communications

This project, developed by Luo-Chenxin, aims to systematically compare the performance of various neural network architectures in radio frequency map prediction tasks. It addresses issues such as high computational complexity and poor generalization of traditional radio frequency map construction methods, providing a standardized evaluation platform for the wireless communication field to support 5G/6G network planning and optimization. The project involves mainstream architectures like CNN, U-Net, and GAN, evaluating model performance using metrics such as MSE and MAE, with application scenarios including network planning and dynamic spectrum management.

In modern wireless communications, traditional radio frequency map construction relies on ray tracing (high computational complexity) and statistical models (poor generalization), which struggle to adapt to the dynamic environments of 5G/6G. Deep learning methods offer advantages such as data-driven approaches, end-to-end prediction, and fast inference. They can learn signal propagation patterns from historical data, reducing reliance on physical models.

This project tests architectures such as CNN (captures local spatial features), U-Net (fuses multi-scale features), GAN (generates high-fidelity maps), and GNN (handles irregular scenarios). Evaluation metrics include MSE, MAE, PSNR, and SSIM. Experiments cover different urban environments, frequency bands, and transmission power configurations, providing a basis for model selection.

This technology can be used in scenarios such as optimizing base station deployment for operators, intelligent spectrum sharing in cognitive radio systems, maintaining communication links for drones/Internet of Vehicles, and indoor positioning and navigation. It reduces field measurement costs and improves communication system efficiency.

Current challenges include high data acquisition costs, insufficient model generalization, black-box characteristics, and high real-time requirements. Future directions may explore hybrid models combining physical knowledge with deep learning, few-shot transfer learning, uncertainty quantification, and lightweight deployment on edge devices.

This project provides a benchmark testing tool for neural network architectures in the wireless communication field, helping researchers select appropriate models. With the development of 6G and AI technologies, radio frequency map prediction will play an important role in scenarios such as intelligent networks and digital twins. Open-source projects promote technical collaboration and progress.