Abstract: Accurate and efficient indoor electromagnetic wave propagation prediction is a fundamental basis for the planning and optimization of wireless communication systems. Traditional methods suffer from significantly increased computational complexity at high frequency bands due to the complex propagation characteristics of electromagnetic waves, making it difficult to meet real-time requirements. Meanwhile, existing deep learning models also face certain limitations in adaptability and prediction accuracy for high frequency bands.This paper proposes a multi-channel U-Net deep learning model (mU-Net), which models antenna directional characteristics, reflective channel contributions, transmissive channel contributions, and free-space path loss as four independent feature channels. By feeding these into a multi-channel U-Net, the network captures the differentiated impacts of various propagation mechanisms. Its asymmetric encoder-decoder structure further integrates multi-source features and resolves complex structural scattering and diffraction details, ultimately producing high-resolution path loss prediction maps.The proposed method deeply fuses high-frequency feature information, overcoming the low real-time performance and insufficient accuracy of existing methods. Finally, path loss prediction results from three different scenarios demonstrate that the proposed mU-Net significantly outperforms typical existing prediction methods in key metrics such as accuracy, structural similarity, and prediction efficiency, with particularly notable performance gains in high-frequency scenarios. This provides a high-accuracy, high-efficiency solution for indoor electromagnetic wave propagation prediction in high-frequency bands.