Function-based and physics-based hybrid modular neural network for radio wave propagation modeling

摘要

A modular neural network approach was used to implement a ray tracing algorithm for radio wave propagation modeling. The goal is to develop a neural network architecture to replace traditional calculations. This method is site-specific so that it can simulate different environments with some acceptable limitation in environment dimensions. In an actual test, the modular neural network is used to predict propagation inside the third floor of the engineering building of CUHK. The average prediction error of the modular neural network is 6.93 dB and 6.01 dB standard deviation for the shadow region, and 5.27 dB with 4.63 dB standard deviation for the line-of-sight region.