Channel Prediction in 6G Non-Terrestrial Networks With Deep Learning: a Physical Layer Analysis

Channel prediction on the user equipment side has recently been identified as one of the main use cases for artificial intelligence in the radio access network. With advanced neural networks, complex channel dynamics can be predicted from current channel estimates, allowing for demodulation reference signals to be transmitted at a lower rate. Reducing reliance on pilot symbols for channel estimation frees valuable resources for data transmission, but this requires highly accurate predictions of future channel states. This paper introduces a lightweight convolutional neural network (CNN)-based encoder-decoder model designed to predict the time-frequency channel response matrix in orthogonal frequency division multiplexing systems. The proposed approach alternates pilot-full and pilot-free slots, leveraging deep learning to predict channel conditions over the latter while maintaining link quality. Through extensive simulations, we assess the model’s effectiveness on key physical layer metrics, with a particular focus on the effective throughput (TP). We show that the proposed predictor achieves TP gains of up to 8% with respect to channel estimation for high modulation orders; on the other hand, low modulation orders’ performance at low energy per bit over noise power spectral density suffer from inaccurate channel predictions. The predictor’s lightweight design, requiring only 160k multiply and accumulate operations and encompassing 5.5k parameters, ensures feasibility for real-time applications in next-generation networks.