A reinforcement learning-driven adaptive hybrid PLC-RF communication architecture for IoT-based smart metering systems

Abstract

With the rapid expansion of smart grid infrastructure, robust and scalable communication is essential to support Advanced Metering Infrastructure (AMI). This paper presents a hybrid communication architecture that combines Power Line Communication (PLC) and Radio Frequency (RF) technologies to enable reliable, adaptive data transmission in smart metering networks. The proposed system employs a PLC-based mesh topology that utilizes existing electrical wiring to relay meter data to a Data Concentrator Unit (DCU), alongside an RF-based star topology that provides direct wireless links between smart meters and the DCU. A cloud-based web application is integrated for real-time visualization of power consumption, network health, and active communication paths. To dynamically select the optimal channel based on real-time latency, packet loss, and Signal-to-Noise Ratio (SNR), a Proximal Policy Optimization (PPO) reinforcement learning agent is implemented and benchmarked against tabular Q-Learning, Double Deep Q-Network (DDQN), and Deep Deterministic Policy Gradient (DDPG). The system is validated through hardware prototyping and subjected to multi-seed training, reward function sensitivity analysis, edge microcontroller profiling, and network-scale simulation to assess deployment viability. Evaluated over 5, 000 deterministic test cases against a reward-derived oracle, the PPO agent achieved an accuracy of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$93.78\%$$\end{document} and an F1 score of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$87.74\%$$\end{document}, attaining the highest single-run performance among all evaluated agents, with statistically significant advantages over Q-Learning and, for F1 score, over DDPG. Multi-seed training across ten initializations confirmed superior convergence stability with an accuracy standard deviation of only \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$8.74\%$$\end{document}, mitigating policy collapse. Edge AI profiling on an Arm Cortex-M4 platform demonstrated that INT8 quantization compresses the model by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.36\times$$\end{document} to 64.7 KB while preserving \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$>99.9\%$$\end{document} of baseline accuracy at a 1.93 ms on-device latency. A reward sensitivity sweep across 16 coefficient perturbations verified policy robustness under varying weight configurations. Finally, a network-scale simulation across 500 to 5, 000 nodes confirmed scale-invariant performance with a Packet Delivery Ratio above \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$92\%$$\end{document} and consistent resilience under node outages, traffic overload, and channel degradation, providing simulation-based evidence of the scalability and resilience of AI-driven adaptive communication for large-scale smart grid deployments.