Energy Efficient Downlink mMIMO Using Dynamic Antenna and Power Adaptation

Massive multiple-input multiple-output (mMIMO) technology and its future evolutions are expected to address the high data rate demands of sixth generation (6G) communication systems. At the same time, network energy savings (NES) is essential in reducing the operational costs and meeting the sustainability goals of network operators. In this regard, we propose a dynamic scheme for joint antenna and power adaptation to improve NES from a user scheduling and resource allocation perspective. Antenna adaptation is performed using the multiple channel state information resource signal (CSI-RS) framework. Furthermore, the recently introduced transmit power-aware link adaptation scheme, referred to as POLITE for short, is used as the power adaptation technique. The proposed scheme adapts to variations in users’ instantaneous traffic and channel conditions to opportunistically maximize NES while also inherently accounting for the user throughput. Numerical simulation results show that the proposed scheme consistently achieves a balance between NES and user perceived throughput (UPT) for different network load conditions. Especially in low and light load conditions, the proposed scheme significantly improves the intra-cell interference and boosts the overall NES, while ensuring that UPT is unaffected.

💡 Research Summary

The paper addresses the dual challenge of meeting the massive data‑rate demands of future 6G networks while reducing the operational energy consumption of massive MIMO (mMIMO) base stations. To this end, the authors propose a dynamic scheme that jointly adapts the number of active antenna ports (spatial domain) and the downlink transmit power spectral density (PSD) (power domain) on a per‑slot basis. Antenna adaptation is realized through the multiple CSI‑RS (channel state information reference signal) framework, where each CSI‑RS configuration corresponds to a predefined set of antenna ports and transceiver chains (TRXs). By switching among these configurations, the gNB can deactivate a subset of its antennas, thereby lowering circuit power and reducing intra‑cell interference when fewer, wider beams are used.

Power adaptation builds on the recently introduced POLITE (Power‑aware Link Adaptation) algorithm. POLITE reduces the PSD by deliberately selecting a lower modulation and coding scheme (MCS) for each user, provided that the resulting data rate remains above a fraction βₙ of the original rate and the target block error rate (BLER) is still satisfied. The factor βₙ is computed in a load‑driven manner: it depends on the user’s buffer occupancy Qₙ, the long‑term average throughput Rₙ,avg, and a scaling constant χ. Consequently, when the network is lightly loaded, βₙ can be set close to 1, allowing modest PSD reduction; under heavy load βₙ is reduced, limiting PSD cuts to preserve throughput.

The authors formalize the problem through two optimization models. The first (P1) is a conventional link‑adaptation problem that maximizes the instantaneous rate Rₙ(k) subject to BLER and PSD constraints for each CSI‑RS configuration m. The second (P2) captures POLITE’s objective: minimize PSD while ensuring the new rate is at least βₙ·Rₙ(k) and the MCS does not exceed the baseline value. Both problems are solved greedily because the wideband CQI reduces the dimensionality of the resource‑allocation task.

Algorithm 1 describes the complete procedure. Starting from the full‑antenna configuration (m = M), the scheduler first computes PF (proportional‑fair) metrics for all users and orders them. It then iterates over the CSI‑RS configurations in reverse order, solving P1 for each m to obtain tentative MCS and required RBs. If the tentative allocation can empty all user buffers, the corresponding m is selected as the active antenna set m′; otherwise the algorithm proceeds to the next configuration. After fixing the antenna set, the POLITE step computes βₙ for each UE, solves P2 to obtain reduced PSD and updated MCS/RB assignments, and finally redistributes any leftover RBs using the PF metric. This loop repeats every transmission slot, ensuring that both CSI and buffer information are continuously exploited.

Performance evaluation employs a 3GPP‑compliant system‑level simulator with realistic traffic models (FTP‑3) and a power‑consumption model that distinguishes micro, light, and deep sleep states of the gNB hardware. Two key performance indicators are used: user‑perceived throughput (UPT), measured as the average application‑layer throughput per FTP‑3 packet, and network energy savings (NES), expressed as the reduction in gNB power consumption relative to a baseline full‑antenna, full‑power scenario.

Simulation results reveal that the proposed joint adaptation scheme achieves a favorable trade‑off between NES and UPT across a wide range of load conditions. In low‑ and medium‑load regimes, the algorithm deactivates a substantial fraction of antenna ports and applies POLITE, leading to PSD reductions of up to 30 % and intra‑cell interference mitigation, while UPT remains virtually unchanged compared to the baseline. Under heavy load, the antenna set cannot be reduced as aggressively, but POLITE alone still yields noticeable power savings (≈10 %) without compromising throughput. The load‑driven βₙ mechanism proves effective at adapting to bursty traffic: users with large buffers receive more RBs, preventing buffer overflow, whereas users with little data are assigned lower MCS, contributing to overall power reduction.

The paper’s contributions are threefold: (1) integration of CSI‑RS‑based antenna selection into the standard 5G‑A/6G signaling framework, (2) adaptation of POLITE’s βₙ parameter to a buffer‑aware, load‑driven formulation, and (3) a practical, low‑complexity algorithm that jointly optimizes spatial and power domains on a per‑slot basis. Limitations include the omission of hardware switching latency and transient power spikes associated with antenna reconfiguration, and the reliance on simulation rather than experimental validation. Future work could incorporate realistic switching models, explore reinforcement‑learning‑based joint optimization of m and βₙ, and test the scheme on a real‑time testbed to quantify implementation overheads.