Downlink power allocation for cell-free massive MIMO systems using unsupervised learning

Cell-free massive MIMO (CF-mMIMO) is a promising wireless technology, outperforming traditional cellular networks in coverage, capacity, and interference management. It involves plenty of geographically distributed access points (APs) connected to a central processing unit (CPU) to serve multiple user equipment (UE) in a coverage area. In this work, the downlink power allocation problem in a CF-mMIMO network is addressed using centralized and distributed deep-learning models, which can learn the complex relationship between large-scale fading (LSF) and power control coefficients. Maximum-ratio (MR) and regularized zero-forcing (RZF) precoding schemes are employed. The deep neural network (DNN) models are trained in an unsupervised learning fashion to maximize the sum of spectral efficiency (sum-SE) while adhering to the per-AP power budget constraint. To achieve this, three custom loss functions are defined, each targeting a specific maximization objective: sum-SE, minimum signal-to-interference-plus-noise (SINR), and product of SINR among the UEs. An LSF-based AP selection algorithm is employed to improve energy efficiency, ensuring that each UE is served by the most contributing APs. Simulation results have proven that the proposed unsupervised learning solutions outperform by up to 20% state-of-the-art supervised counterpart in terms of spectral efficiency, and by more than 22% in terms of energy efficiency. Moreover, the unsupervised learning method circumvents the arduous task of label generation, as required by supervised learning techniques.