Introduction

The rapid deployment of high-end power systems has driven demand for intelligent and efficient energy management technologies, with Advanced Metering Infrastructure (AMI) emerging as a cornerstone of smart grid development. AMI enables utilities to perform real-time monitoring, control, and analysis of electricity usage. In this context, “real-time” refers to data collection at much higher frequencies (e.g., minutely or hourly intervals) than traditional monthly manual readings, resulting in enhanced operational efficiency, reduced energy losses, and improved demand response programs1. At the core of AMI lie smart meters, which serve as intelligent endpoints capable of capturing fine-grained consumption data and enabling two-way communication between consumers and utility providers. Recent reviews on smart meter data analytics, as reviewed in2, emphasize the critical role of leveraging massive volumes of meter data through descriptive, predictive, and prescriptive analytics to enhance grid efficiency, sustainability, and decision-making. From an Internet of Things (IoT) perspective, AMI can be viewed as a large-scale vertical IoT deployment, where smart meters act as resource-constrained edge devices connected through heterogeneous communication technologies to utility back-end and cloud services. A fundamental requirement for AMI’s success is a scalable and dependable communication network capable of interconnecting large numbers of geographically dispersed smart meters. Traditional solutions, Power-Line Communication (PLC) and Radio-Frequency (RF), each exhibit critical limitations: PLC suffers from noise and attenuation3, while RF is prone to interference and coverage gaps4. PLC leverages existing electrical wiring for data communication, offering cost-effective deployment, but endures significant signal attenuation, noise, impedance mismatches, and electromagnetic interference especially in legacy grids3. Galli et al.3 outlined the role of PLC in both in-home automation and utility-side grid monitoring, showing how it can enable widespread coverage without the need for additional cabling. However, PLC faces challenges including impulse noise, frequency-selective fading, and high channel attenuation due to varying electrical loads on the power lines. Challenges like impedance mismatches in PLC are addressed in recent tutorials on modem coupling circuits5, which highlight design trade-offs and safety considerations while providing insights into enhancing signal quality over low-voltage lines. Conversely, RF communication provides significant flexibility and ease of deployment. RF networks often utilize star topologies and are better suited for long-distance communication compared to PLC6. However, its performance can be adversely affected by external factors such as electromagnetic interference, physical obstructions, and limited coverage in densely populated or complex environments, as highlighted in4. To address these individual shortcomings and ensure robust coverage, hybrid approaches combining PLC and RF have gained significant attention. Avdaković et al.7 reviewed smart metering communication technologies across HAN, NAN, FAN, and WAN, highlighting the need for scalable, secure, and interoperable systems in low-voltage networks, which motivates hybrid PLC–RF solutions. Pittolo et al.8 showed that dynamic switching in hybrid PLC/Wi-Fi systems enhances robustness and throughput, supporting real-time adaptation in smart grids. Similarly, multi-interface nodes that switch between PLC and RF based on link quality provide resilient communication, while approaches like AAV-5G9 demonstrate that integrated wired–wireless architectures improve coverage and self-healing. These findings inform the proposed design of a PLC mesh combined with an RF star topology for reliable AMI data transmission.

Related work

Two notable hybrid communication approaches in smart grids are G3-Hybrid10 and PRIME Hybrid11, both enhancing reliability and coverage by enabling seamless switching between PLC and RF. G3-Hybrid extends G3-PLC by unifying PLC and RF under a common MAC layer, selecting the channel per frame based on metrics like SNR, packet error rate, and routing cost, with retries over the alternate medium and deduplication to prevent redundancy. PRIME Hybrid integrates RF and PLC at the PHY layer with a shared MAC, dynamically selecting the medium using link quality indicators and thresholds, often leveraging IEEE 802.15.4 and CSMA/CA, while maintaining backward compatibility with PLC-only nodes12. Recent developments, such as those reviewed by Avdaković et al.7, highlight hybrid PLC/RF systems for smart grid monitoring, achieving up to 99% coverage in urban settings through interdisciplinary power and information engineering approaches. Similarly, Huc et al.13 provide a technical overview of G3-PLC Hybrid, demonstrating improved data rates and interference mitigation in field trials.

Early channel selection in hybrid systems often relied on threshold-based heuristics, which provide low-complexity switching without requiring data-driven training. For example, Bithas et al.14 propose an improved threshold scheme for wireless systems, applicable to PLC/RF hybrids, where channels are selected based on SNR or error rates falling below optimized thresholds. Using Markov chain analysis, it reduces complexity and outage probability by 20-30% compared to continuous monitoring, though it lacks adaptability to non-identical channel distributions common in dynamic grids. Critically, even adaptive threshold methods share a fundamental structural limitation: they optimize a single scalar trigger condition (e.g., SNR < \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma _{\text {th}}$$\end{document}) for each metric independently, and cannot simultaneously arbitrate among multiple correlated objectives such as packet loss, latency, and SNR under time-varying channel conditions. Furthermore, threshold schemes must be statically configured at deployment time and require expert re-calibration whenever traffic patterns, topology, or channel characteristics change, making them poorly suited to the heterogeneous and non-stationary PLC and RF environments observed in operational AMI networks.

Supervised learning (e.g., SVM) and semi-supervised approaches have been used to classify link quality, operating effectively when partial labels are available. However, training robust supervised models typically requires substantial labeled datasets, limiting their immediate plug-and-play use in hybrid PLC–RF systems due to the scarcity of real-time metric data like latency, packet loss, and SNR across diverse grid topologies. Moussa et al.15 compare ML models, including SVM with linear and non-linear kernels, for predicting PLC link availability in smart grids, achieving 85-86% accuracy on a 1,000-reading dataset of SNR, RSSI, and CINR. While these methods outperform simpler heuristics in noisy environments, the study highlights data correlation issues and the need for RF fallback, underscoring the challenges of applying purely supervised methods in highly variable scenarios without continuous retraining.

Beyond heuristic and supervised-learning-based schemes, recent work has explored the use of advanced routing and intelligence in AMI communication protocols. For instance, fault-tolerant routing schemes have been proposed to ensure reliable data aggregation in smart grid meshes16. Furthermore, reinforcement learning has been integrated directly into AMI routing, with a notable example being Q-RPL17, which integrates a Q-learning mechanism into the RPL routing protocol to optimize parent selection in wireless AMI mesh networks. By learning the next-hop choice based on link quality and routing performance, Q-RPL improves packet delivery ratio (PDR) by approximately 5–16% compared to standard RPL, RPL+, and ML-RPL variants, while achieving median end-to-end delays in the range of 177–286 ms and compliant factors of 94–100% in realistic smart meter deployments. This demonstrates that reinforcement learning can effectively enhance reliability and latency at the network (routing) layer of smart grid communication infrastructures. Unlike Q-RPL, which applies Q-learning at the routing layer to optimize parent selection in wireless AMI meshes, the proposed approach targets a complementary problem: node-level hybrid media selection between PLC and RF interfaces based on real-time link metrics. Rather than modifying the routing protocol, the proposed PPO agent operates as an interface-selection layer that learns when to use each medium while treating the underlying PLC and RF stacks as black boxes.

Reinforcement Learning (RL) offers a model-free alternative, learning policies through environmental interaction. As Gómez-Rodríguez et al.18 indicate, RL suits smart grids with unpredictable dynamics. A comprehensive review by Kim et al.19 on smart grid evolution and RL applications addresses challenges in power flow optimization and demand response, showing RL’s potential for scalable grid management. Similarly , a survey in20 regarding deep RL for smart grid operations highlights improvements in energy scheduling and fault resilience. For fault detection, Olojede et al.21 apply hybrid ML-RL models for predictive maintenance, achieving 88-93% accuracy in insulation fault detection. In real-time optimization for microgrids, Liu et al.22 use DDPG, reducing operating costs by up to 30% under uncertainties from renewables and loads. While Q-Learning suits discrete problems, extensions like DQN, DDQN, and DDPG enhance scalability, though PPO23 balances exploration and convergence, making it ideal for the dynamic hybrid networks.

More recently, the convergence of edge AI and IoT communication has produced several relevant research directions. Federated learning has been explored for privacy-preserving smart meter analytics, enabling distributed model training across AMI nodes without centralizing sensitive consumption data24. In the domain of lightweight embedded AI, TinyML frameworks have demonstrated the feasibility of deploying reinforcement learning agents on microcontrollers with as little as 32–512 KB of RAM through aggressive model compression techniques such as INT8 quantization, pruning, and knowledge distillation25. For hybrid communication reliability, recent analyses of smart grid communication infrastructure emphasize the importance of selecting optimal technologies across HAN, NAN, FAN, and WAN layers to ensure reliable, interoperable, and scalable data exchange7. These developments collectively inform the design choices in this work and highlight the growing maturity of AI-driven communication management in smart grid infrastructure.

The literature shows a clear trade-off: non-AI heuristic methods14 are simple but static, lacking the adaptability for dynamic grid conditions. Conversely, supervised ML models15 are data-driven but require large, pre-labeled datasets that are impractical for this application. This creates a research gap for a solution that is both adaptive and model-free. To summarize the trade-offs discussed, Table 1 provides a qualitative comparison of the different channel selection methodologies. This comparison highlights the gap addressed by this work, justifying the use of an RL-based approach over simpler, non-AI baselines.

Approach | Application | Adaptability | Data requirement | Complexity | Key limitation
Heuristic14 | Wireless channel selection | Static (Fixed Rules) | Low (Manual Calibration) | Very Low | Fails in changing network conditions
Supervised & Semi-Supervised15 | PLC node availability | Static (Post-Training) | High (Pre-Labeled) | Medium | Challenges obtaining diverse labeled datasets
Q-RPL17 | AMI routing (network layer) | Dynamic (Learns) | Low (no pre-labeled) | Medium-High | Operates at routing layer, not physical
RL (This Work) | PLC/RF interface selection | Dynamic (Learns) | Very low | High | Advantage: Addresses data limitations

Contribution

This paper proposes a hybrid PLC–RF communication system for smart metering networks, which dynamically selects the optimal communication channel based on real-time network conditions. The system integrates a PLC mesh topology, where smart meters relay data through neighboring nodes to a Data Concentrator Unit (DCU), providing enhanced coverage and fault tolerance. Simultaneously, an RF star topology implemented via the TI 802.15.4 stack26 that enables meters to transmit directly to an RF-enabled DCU, which subsequently forwards data to a cloud server for advanced processing and visualization.

While existing standards like G3-Hybrid and PRIME Hybrid provide mature dynamic switching capabilities at the MAC/PHY layers, the primary novelty and contribution of this work lies in the application and comparative analysis of four RL models (PPO, DDQN, DDPG, and Q-Learning) for intelligent interface selection. This work is distinct from existing approaches in the following respects. First, unlike G3-Hybrid and PRIME Hybrid, which perform threshold-based switching at the MAC/PHY layer using fixed link quality indicators, the proposed approach employs a learned policy that adaptively weights multiple channel metrics simultaneously without requiring manual threshold calibration. Second, unlike Q-RPL17, which applies Q-learning at the routing (network) layer to optimize parent selection in wireless meshes, this work targets a complementary and orthogonal problem: physical-layer interface selection between heterogeneous PLC and RF media. Third, unlike supervised learning approaches such as15, which require pre-labeled datasets and periodic retraining, the RL-based approach is model-free and learns directly from environmental interaction, making it suitable for deployments where labeled communication data is scarce. Instead of relying on static thresholds, the proposed PPO agent acts as an adaptive interface-selection layer that continuously evaluates real-time network performance metrics such as packet loss, latency, and signal-to-noise ratio (SNR) to dynamically decide the optimal medium. The models are trained in a novel simulation environment that is calibrated against empirical data from a physical hardware testbed. To validate real-world deployment feasibility, policy stability, and scalability, this work also introduces: (1) a multi-seed validation across ten distinct training seeds, verifying PPO’s superior convergence consistency and low variance (8.74% standard deviation in testing accuracy) against alternative off-policy algorithms; (2) edge hardware deployment profiling on an Arm Cortex-M4 microcontroller platform, where full INT8 quantization achieves a 2.36× footprint compression (64.7 KB Flash occupancy) and an on-device latency of 1.93 ms while retaining >99.9% of the baseline accuracy; (3) a reward function sensitivity analysis under 16 parameter perturbations (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 25%$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 50%$$\end{document} sweeps) showing PPO’s robustness to varying latency, loss, and switching cost weights; and (4) a simulation-based network-scale scalability and stress-testing evaluation across topologies of 500 to 5, 000 communication nodes, demonstrating scale-invariant Packet Delivery Ratio (PDR) above 92%, low centralized batch latencies of 2.27 to 22.9 _μ_s, and strong resilience under node outages, traffic overload, channel degradation, and impulsive noise.