System architecture

Fig. 1: Hybrid communication system architecture.

The architecture in Fig. 1 enables reliable communication from residential smart meters to the utility provider under diverse channel conditions. This architecture reflects a realistic Advanced Metering Infrastructure (AMI) deployment, structured around four key components: end-user sensor nodes (smart meters), Data Concentrator Units (DCUs), the utility’s Head-End System (HES), and the cloud-based web application. Together, these components form a hybrid communication pipeline that supports real-time data collection, aggregation, management, and centralized analytics. Although quantitative scalability limits depend on the deployment environment, standard configurations support multiple PLC hops and extended RF ranges under typical loads, establishing a robust framework for large-scale AMI. Each end-user sensor node is installed at the consumer’s premises and serves as the first point of data generation. These nodes are equipped with a dual-mode communication system that supports both power line communication (PLC) and sub-GHz radio frequency (RF) transmission. The end-user nodes and DCU components in the proposed architecture leverage principles from model-based safe RL for distribution systems27, which demonstrate how reinforcement learning can be constrained to guarantee operational safety while optimizing system performance. By analogy, the proposed PPO-based framework dynamically switches between PLC and RF communication paths, ensuring consistent data transmission under varying interference and signal conditions while avoiding unsafe or unstable routing decisions. The hardware stack comprises an STM32F401RCT6 microcontroller for onboard data processing and control operations28, a TI1352B1 RF module operating on the TI 15.4 software stack for IEEE 802.15.4g-compliant wireless communication29, and an LX200V50 PLC module for data transfer over the electrical grid infrastructure30. Additionally, each node is embedded with a lightweight AI agent based on the Proximal Policy Optimization (PPO) reinforcement learning model. This agent dynamically selects the most reliable communication path in real time by evaluating environmental metrics such as signal-to-noise ratio (SNR), packet loss, and latency. The hybrid design not only enhances communication reliability but also improves system continuity by allowing intelligent switching between RF and PLC depending on link quality.

Data generated by the sensor nodes after getting it from the smart meters at homes is transmitted through two primary communication pathways: PLC and RF. In the PLC mode, data is sent over the existing electrical wiring to a nearby Data Concentrator Unit (DCU). While PLC is cost-effective and utilizes pre-existing infrastructure, it is susceptible to noise, impedance mismatches, and signal attenuation especially in legacy electrical environments. These challenges necessitate the use of effective routing protocols and error correction techniques. Conversely, the RF path relies on low-power wireless transmission using the 802.15.4g standard. In this mode, multi-hop communication is employed, where data is relayed through intermediate nodes until it reaches the Data Concentrator Unit (DCU). This wireless channel acts as a redundant path, ensuring communication continuity in scenarios where PLC links are degraded or interrupted. To optimize performance under changing network conditions, an adaptive channel selection strategy is implemented using the PPO model embedded in each sensor node. This model continuously monitors real-time metrics and autonomously selects the most suitable channel for data transmission. Compared to traditional reinforcement learning approaches, including Q-Learning or DDPG, PPO demonstrates more stable convergence, improved exploration–exploitation balance, and greater robustness in highly variable environments . The Data Concentrator Units (DCUs) aggregate data from end-user sensor nodes within their area. These DCUs receive data via both RF and PLC links, buffer the information, and then utilize a high-bandwidth backhaul (such as cellular or Ethernet) to transmit the aggregated data to the utility’s central Head-End System (HES). Strategic placement of these DCUs, often downstream from a distribution transformer, ensures full coverage for a specific neighborhood or low-voltage segment, guaranteeing consistent data flow. All collected data from across the AMI network converges at the HES, which handles data validation, storage, and management. The proposed cloud-based web application then acts as the central intelligence and visualization layer by interfacing with the HES. This platform pulls the managed data to provide dashboards for both end users and utility administrators. Users can view real-time energy consumption and historical trends, while administrators are given access to advanced system controls, anomaly detection tools, and maintenance alerts. The web application backend is designed for deployment on a secure cloud infrastructure and supports explicit role-based access control (RBAC), billing analytics, and fault management features. This RBAC implementation ensures end-users see only their own data, while utility administrators access system-wide diagnostics and controls. A comprehensive security analysis is essential in AMI deployments to mitigate risks such as PLC eavesdropping, potential RF jamming attacks, and adversarial perturbations targeting the reinforcement learning agent’s observation space. While a full exploration of these cryptographic and adversarial robustness requirements is beyond this paper’s immediate scope, the proposed architecture assumes standard security protocols (e.g., TLS/SSL for backhaul, AES encryption at the physical/MAC layers) are enforced, and future extensions will investigate adversarial training to fortify the PPO agent in real-world smart grid contexts.