ML-based agglomerative hierarchical clustering and MMSE precoding for fair user scheduling in LEO MU-MIMO systems

This paper addresses the problem of user scheduling in Low Earth Orbit (LEO) High Throughput Satellite (HTS) multi-user (MU) Multiple-Input-Multiple-Output (MIMO) systems. We propose a Machine Learning (ML)-based hierarchical clustering approach i.e. AGNES (Agglomerative Nesting), where the hierarchy of clusters is established by splitting and merging users across different levels in the service area. The primary objective is to minimize inter-beam interference caused by aggressive frequency reuse in the feed space (FS) of LEO-HTS. The proposed scheduling algorithm is computationally efficient, thus offering a novel solution that maximizes the number of users accommodated within minimal time slots. Additionally, it provides flexibility to enhance performance by increasing the number of time slots and relaxing cluster size constraints. Three power normalization methods are used to apply a per-cluster Minimum Mean Square Error (MMSE) beamforming matrix. The algorithm's performance is compared against two benchmark schedulers: (1) a position-based random scheduler that generates a beam lattice and randomly selects one user per beam for clustering, and (2) a graph-based scheduler that uses the computationally demanding maximum clique technique to improve graph density. Extensive simulation results demonstrate that the proposed scheduler significantly outperforms both the position-based and graph-based approaches in terms of efficiency and performance. Moreover, it eliminates the need for hard threshold optimization across various power normalization schemes.