BEACON: Beam Prediction with Efficiency for Advanced V2I Communication Networks

Millimeter-wave (mmWave) and terahertz (THz) frequencies enable high data rate applications in modern wireless communication, with vehicle-to-everything (V2X) communication being critical for traffic management, safety, and autonomous driving. Beam prediction in highly mobile V2X scenarios seeks to reduce the significant overhead of traditional beam training in mmWave systems. Prior studies on beam prediction are often limited to current-time beam selection or rely solely on synthetic datasets, lacking real-world validation. This study proposes a Liquid Time Constant Neural Network (LTC-NN)-based framework for future beam prediction, leveraging its adaptive time constants to handle dynamic vehicular data. We evaluate the approach on the real-world DeepSense 6G and simulated DeepMIMO datasets using lookback lengths ℓ∈{5,8} for prediction horizons 𝜏∈{1,3}. Averaged across all scenarios and settings, LTC-NN achieves a top-1 accuracy of 48.65% versus 44.23% for LSTM and reduces average power loss to 0.309 dB compared to 0.447 dB for LSTM. At a 90% reliability threshold, beam training overhead savings exceed 94.4% for LTC-NN, compared to 93.6% for LSTM. These results highlight the novel application of LTC-NN for V2I beam prediction, addressing a critical gap in efficient, accurate beam management. This research offers a practical approach to enhance mmWave V2I communication systems, paving the way for robust, energy-efficient solutions in dynamic vehicular environments.