Unsupervised Learning-based State-Space Model for Predicting Time-varying Channels

Accurate channel prediction is critical to track the large channel variations caused by high user mobility. For this prediction task of complex-valued time series data, the state-of-the-art solutions employed the model-based state-space model with limited expressive capacity, or the data-driven Transformer that lacks interpretability. Recent works attempted to alleviate both drawbacks by fusing neural networks into the filtering workflow of state-space model. However, this inspiring work was not compatible with channel prediction in realistic settings. Specifically, it was intended for the real-valued low-dimensional filtering task, and it necessitates labels for training. What's worse, the adopted neural network is too overparameterized and thus data hungry. To resolve the above three issues in practice, this paper proposes a novel hybrid state-space model to predict complex-valued high-dimensional wireless channels, which removes the necessity of labels and owns an increased data efficiency. Numerical results demonstrate its superior performance in prediction accuracy compared to the state-of-the-art model-based and data-driven methods. Its great scalability in dimension and robustness to high mobility make itself a promising candidate for deployment on resource-limited edge devices.