A Dynamic Power Allocation Approach for Downlink Cell-Free Massive MIMO With Graph Neural Network

Cell-free (CF) massive multiple-input multiple-output (MIMO) is regarded as a potentially transformative paradigm in 6G Internet of Things (IoT) networks. The efficient downlink power allocation is crucial in CF massive MIMO, especially in the presence of IoT device mobility. In this study, we propose a novel approach based on a dynamically connected graph neural network (DCGNN) to optimize the downlink power allocation. The effective signal-to-interference-and-noise ratio (SINR) of the IoT devices is modeled in a dynamically connected graph, where the nodes represent the access point (AP)-IoT device links, the edges represent the dominant interference between different nodes. The dominant interference is selected by a sorting function in each time slot. By leveraging the graph topology, the proposed DCGNN-based approach can effectively exploit the SINR knowledge in the interference circumstances. We study the maximization of sum-spectral efficiency (SE) downlink power allocation in CF massive MIMO with correlated channel fading. The numerical results show that the proposed DCGNN-based approach outperforms the state-of-the-art approaches on the sum-SE performance. More specifically, the sum-SE achieved by the proposed DCGNN-based approach is higher than the deep reinforcement learning-based approaches and other benchmarks.