Deep Learning Based Uplink Power Allocation in Multi-Radio Dual Connectivity Heterogeneous Wireless Networks

Multi-Radio dual connectivity (MR-DC) has been emerged as a promising 3rd Generation Partnership Project (3GPP) feature aiming at improving per-user data throughput enabling traffic aggregation from two base stations (BSs) simultaneously operating the same or different radio access technologies (RATs). In MR-DC uplink (UL) data transmission, UL power assigned to the two interfaces may exceed the maximum transmission power of the user equipment (UE) leading to increased UL effective interference at the network. An optimal power allocation strategy can significantly reduce the effective UL interference which also leads to the enhancement of network throughput. But in the MR-DC scenario, dynamic switching of node assignment and consideration of effective UL transmission in a given sub-frame introduce additional challenges, making existing methods like semi-static or iterative and heuristic optimization algorithms inefficient. In this study, we aim to design a deep convolutional neural network (DCNN) model, more specifically a U-Net based model, for intelligent power allocation in the MR-DC network. A constrained optimization function is formulated considering data rate maximization, UL effective interference reduction, and effective UL transmission towards any node in the MR-DC, as well as an Interior-point method (IPM) algorithm is utilized to generate a labeled dataset. The proposed model is trained to predict the optimal power label more efficiently, especially when dealing with varying numbers of UEs and BSs. Finally, we have depicted that the proposed U-Net based model can achieve up to 10.98% total throughput improvement predicting optimal power label for 88 UEs within 0.07 seconds through simulations.