In recent years, with the rapid development of optical neural networks, all-optical diffractive deep neural networks have been applied across various fields. However, there is a mismatch between the numerical modeling and the physical deployment of all-optical diffractive deep neural networks. This discrepancy arises because the input signal is not an ideal plane wave, as the 300 GHz THz signal is transmitted through a horn antenna to ensure sufficient light intensity. To address this issue, we propose an antenna aperture compensation algorithm for all-optical diffractive deep neural networks. Classification experiments on the MNIST handwritten digit dataset validated the effectiveness of this approach, demonstrating the reliability of the compensation algorithm and its potential to improve the overall performance of optical systems in future experiments.