Model-Agnostic Channel Prediction with Meta Predictive Recurrent Neural Networks

Vehicle-to-vehicle (V2V) communications are crucial for intelligent vehicle control, reducing accidents, and increasing energy efficiency. This paper focuses on accurate and generalizable V2V channel prediction, which is a critical component of V2V communications systems, but which is challenging due to the complexity of different V2V environments. Addressing this, we leverage recent machine learning (ML) advancements for time series modeling, applying innovative model-agnostic ML algorithms to this complex issue. Our key innovation is the creation of a unique Squeeze and Excitation long-short-term memory (SE-LSTM) recurrent neural network. This model combines a squeeze and excitation block, global information embedding, and stacked memories within the LSTM framework. It excels in data modeling across diverse domains (Time, Doppler, Delay, and Angular) and effectively captures both intra- and inter- sequence dependencies in Channel State Information (CSI) data. We also utilize meta Pseudo-Label learning for network training, enhancing generalization across different geographic layouts. Extensive evaluations using data from our measurement campaign show that our method surpasses both baseline and state-of-the-art predictive algorithms, with over 53.4%/4.9% improvement in same-geometry tests and 25.4%/9.5% in cross-geometry tests.