Discussion

To contextualize the performance of the proposed PPO-based hybrid PLC/RF channel selection, a comparison is made with representative recent AI-based approaches in smart grid communication and operation. Table 17 summarizes the main application domain, AI method, and reported quantitative performance for each work.

Work | Application | AI Method | Metric | Performance
This work (PPO) | Hybrid PLC/RFchannel selection | PPO-based RL | Accuracy / F1 score | 93.78% / 87.74%
Avg. latency / loss | 4.94 ms / 4.07%
Q-RPL17 | AMI wireless meshrouting (network layer) | Q-learning-basedrouting | PDR improvement | +5–16% vs RPL variants
Median end-to-end delay | 177–286 ms
Compliant factor | 94–100%
Moussa et al.15 | PLC node availabilityprediction | SVM, AdaBoost,Random Forest | Accuracy / F1 score | 87% / 0.87 (AdaBoost)
Precision / recall | 0.90 / 0.86
Liu et al.22 | Microgrid economic scheduling | DDPG | Cost reduction vs baselines | 29-31% vs DQN;17-18% vs SAC
Olojede et al.21 | Grid fault detection | Hybrid ML/RL | Fault classification accuracy | 88–93%

As shown in Table 17, the proposed PPO-based media selection model achieves 93.78% accuracy and an F1 score of 87.74% (evaluated against a reward-derived oracle on a single fixed-seed run, as described in “Reinforcement learning for channel selection” section), comparable to or slightly higher than other AI-based methods in smart grid contexts, such as the AdaBoost classifier in15 (87% accuracy, F1 of 0.87). Q-RPL17 demonstrates that reinforcement learning can significantly improve routing performance at the network layer, achieving a 5-16% increase in packet delivery ratio and high compliant factors (94-100%) in AMI wireless meshes, while the present work focuses on a complementary problem: selecting between heterogeneous PLC and RF links at each smart meter based on real-time link metrics. Within this broader landscape, where Liu et al.22 and Olojede et al.21 report strong gains in economic scheduling and fault detection accuracy using deep RL and hybrid ML/RL approaches, the proposed PPO-based framework extends the use of reinforcement learning to hybrid PLC-RF AMI communication, achieving high decision accuracy and low latency/loss without relying on large labeled datasets and operating directly on link-quality measurements.

Scalability and deployment considerations

It is important to note that the hardware prototype comprised three RF sensor nodes and five PLC nodes, and serves as a proof-of-concept demonstration of communication feasibility and RL-based channel selection under laboratory conditions. The scalability analysis reported below is derived entirely from simulation and should not be interpreted as validated deployment performance at scale. The strong testing results of the PPO agent, achieving 93.78% accuracy (Table 8) and the lowest average packet loss (Table 9), indicate that the learned policy is suitable for real-time deployment. This is further corroborated by the simulation-based network-scale evaluation in “Simulation-based network scalability and stress testing” section, where in simulation the PPO policy maintained a PDR above 92% across topologies of 500 to 5,000 nodes with scale-invariant performance (Table 14). The deterministic on-device inference latency of 1.93 ms on an STM32 Cortex-M4 microcontroller (Table 12) is negligible relative to link-layer transmission times, confirming computational feasibility at the edge. Alternatively, as shown in Table 16, inference can be centralized at the DCU level, yielding an amortized per-node latency of 2.27–22.9 _μ_s without introducing computational bottlenecks. The main overhead is the dual-interface design, but this additional hardware cost is offset by the gain in resilience, as the hybrid PLC–RF architecture maintains connectivity in conditions where single-medium solutions are likely to fail. In terms of practical deployment cost, the dual-interface design adds the bill-of-materials cost of a second communication chipset, the LX200V50 PLC modem alongside the CC1352R RF transceiver, relative to a single-medium meter design, but removes the need for a separate cellular modem and its associated recurring data subscription fees, which is a common alternative AMI backhaul option. A detailed bill-of-materials and total cost of ownership comparison across these deployment options was outside the scope of the present study. Regarding interoperability with established AMI and grid standards, the RF interface already implements IEEE 802.15.4g at the physical and MAC layers through the TI 15.4 stack, as described in “RF network” section, so interoperability at that layer is already achieved by construction. DLMS/COSEM is the dominant international standard for smart meter data modeling and application-layer exchange, and adopting it at the Head-End System interface, in place of the current proprietary REST API and MongoDB-based backend, would allow the system to exchange meter data with third-party utility head-end systems using a standardized object model, though this has not been implemented in the current prototype. IEC 61850 is primarily used for substation automation and protection communication rather than meter-level AMI data exchange, and would become relevant mainly if the DCU to utility integration in the present architecture were extended toward broader substation automation; within the AMI-focused scope of the present work, it is not directly applicable. By reusing existing power lines and unlicensed sub-GHz spectrum instead of relying solely on cellular infrastructure, the proposed approach supports scalable AMI roll-outs with predictable long-term operational costs and aligns with the need for interoperable communication solutions in low-voltage distribution networks. Furthermore, policy robustness under reward weight variations (“Reward function sensitivity and policy robustness analysis” section) minimizes the need for high-precision coefficient tuning, which is a major deployment benefit when channel distributions drift over time.

It is important to note that the current hardware validation was conducted at laboratory scale: RF testing occurred at distances of 0.5–1.0 m, and PLC mesh nodes were spaced 1.2–2.0 m apart on the same electrical phase. A subsequent point-to-point PLC test achieved a range of approximately 9 m in a separate laboratory environment, confirming that connectivity is maintained provided the modules remain on the same electrical phase. These distances are representative of a proof-of-concept demonstration rather than an operational AMI deployment. However, the communication modules employed offer substantially greater range under field conditions. The TI CC1352R RF module, operating at sub-1 GHz frequencies with the TI 15.4 stack, supports line-of-sight ranges exceeding 1 km at 10 dBm transmit power29, and the LX200V50 PLC module is rated for communication distances up to 500 m over low-voltage power lines30. A link-budget analysis using the measured RSSI of approximately -40 dBm at 10 dBm transmit power (Fig. 3) and assuming a receiver sensitivity of -110 dBm yields a fade margin of approximately 70 dB, which is sufficient to accommodate path loss at distances of several hundred meters in typical urban environments. In a scaled deployment, the hierarchical architecture (sensor nodes → DCU → HES) inherently supports geographic scaling, as each DCU covers a neighborhood segment. The per-node PPO agent operates independently, requiring no inter-node coordination for channel selection, which eliminates communication overhead that would otherwise grow with network size. Routing complexity within the PLC mesh is managed by the VPN overlay and standard mesh routing protocols, and scaling to larger topologies would primarily involve adding additional DCUs rather than modifying the channel selection mechanism itself. Nonetheless, comprehensive field trials at representative AMI distances and node densities remain essential future work to validate the system under real-world propagation conditions, varying traffic loads, interference scenarios, and failure modes.

Regarding performance over longer feeders specifically, low-voltage distribution feeders extending hundreds of meters introduce frequency-selective fading and higher, more variable attenuation driven by impedance mismatches and branch topology, effects not explicitly modeled by the calibrated latency, packet loss, and SNR ranges in Table 3 or Table 4, as already noted among the limitations of the present simulation environment. Because the interface-selection mechanism observes channel quality only through these three statistical indicators rather than through the underlying physical cause of degradation, its expected response to a longer, higher-impedance feeder is a shift in the observed PLC latency, packet loss, and SNR distribution rather than a qualitatively different failure mode for the architecture itself. The closest existing evidence for this response is the channel degradation stress scenario in “Simulation-based network scalability and stress testing” section, where halving PLC SNR and doubling RF packet loss caused the PPO agent PLC selection ratio to fall from 48.1 percent to 34.4 percent, demonstrating that the learned policy shifts toward increased RF reliance as PLC channel quality deteriorates. Since RF star connectivity to a nearby DCU is largely insensitive to PLC feeder length, an analogous or stronger shift toward RF would be expected over long, high-impedance feeders in practice. Explicit validation over real distribution feeders with representative impedance profiles and impulsive noise statistics, however, has not been performed. Regarding RF specific propagation environments, the experiments reported in “RF network” section were conducted under controlled, single path indoor laboratory conditions at short range, and do not include urban multipath fading, industrial electromagnetic interference, non line of sight obstruction, or the RF co-channel interference that would arise among densely co-located smart meters sharing the same sub-GHz spectrum. As already noted in “Simulated training environment” section, the calibrated simulation environment likewise does not model multipath fading or co-channel RF interference explicitly, so the empirically observed SNR and packet loss ranges in Table 3 reflect this benign laboratory propagation environment rather than these harsher conditions. Node density effects are evaluated only at the level of the Monte Carlo network-scale simulation in “Simulation-based network scalability and stress testing” section, which models per-DCU traffic congestion and correlated impulsive noise but does not model RF physical-layer contention or interference among densely co-located transceivers. Because the interface-selection mechanism responds to any combination of degraded latency, packet loss, and SNR regardless of the underlying propagation cause, it is expected to remain functionally applicable under these conditions, but quantifying its performance under urban multipath, industrial interference, non line of sight links, and dense co-channel deployments has not been performed.

Security considerations

Security is a critical requirement for AMI deployments operating over critical infrastructure. The proposed architecture incorporates multiple layers of security through the underlying communication stacks. The RF communication layer, implemented via the TI 15.4 stack on the CC1352R module, provides MAC-level security through AES-128 CCM* encryption and authentication as specified by the IEEE 802.15.4 standard26. The CC1352R hardware includes a dedicated cryptographic accelerator supporting AES 128- and 256-bit operations, which offloads encryption from the main processor and enables efficient key management with minimal computational overhead. For the PLC backhaul, the VPN overlay provides encrypted tunneling between mesh nodes. The cloud-based backend enforces TLS/SSL for all data in transit and implements role-based access control (RBAC) to restrict data visibility.

From a machine learning security perspective, the PPO agent’s observation space (latency, packet loss, SNR) could be vulnerable to adversarial perturbation through metric manipulation, where a malicious actor spoofs or jams channel measurements to bias the agent’s switching decisions. Potential attack vectors include RF jamming to artificially degrade the RF channel metrics, or injection of noise on the power line to manipulate PLC SNR readings. Mitigating such threats would require anomaly detection on the input observation space (e.g., statistical bounds checking, temporal consistency verification) and adversarial training to improve the agent’s robustness to perturbed inputs. These defensive mechanisms, along with a comprehensive threat model covering authentication, encryption overhead, and privacy implications of meter-level communication data, constitute important directions for future investigation.

Beyond adversarial perturbation of the observation space discussed above, several additional AI specific threat categories are relevant to an RL enabled communication architecture and are discussed here in relation to the specific design of the proposed system. Reward poisoning, in which an attacker corrupts the signal used to update the policy, is not a live threat against the currently deployed agent because, as established in “Edge deployment feasibility and hardware profiling” section, the deployed PPO policy is trained offline in the calibrated simulation environment and flashed as a fixed policy rather than updated online from field observations. The attack surface for reward poisoning would only become relevant if a future online or continual learning extension were introduced, at which point field derived reward signals would need to be authenticated and bounded before being used for policy updates. Model extraction is a concern specific to edge deployed machine learning, since the 64.7 KB INT8 policy reported in Table 12 is physically stored in the Flash memory of each smart meter node, some of which may be installed in physically accessible locations. Without protection, an attacker with physical access could attempt to read out the Flash contents through the microcontroller debug interface and recover the trained policy. The STM32F401RCT6 supports hardware Flash readout protection, which can be configured to disable debug port access to the Flash memory and mitigate this class of attack, and its use is recommended for field deployment though not yet enabled in the current laboratory prototype. Replay attacks, where previously captured valid frames are retransmitted to manipulate node state or channel metric readings, are addressed at the RF layer by the frame counter mechanism defined within the IEEE 802.15.4 CCM* security suite already employed by the TI 15.4 stack, which the current prototype assumes is enabled as part of the standard MAC layer security configuration described above, though explicit validation of replay resistance on the physical hardware has not been performed. False data injection targeting the meter consumption data path, as distinct from the channel metric observations discussed above, is mitigated at the application layer through the TLS protected backhaul and the role based access control already implemented in the web application, which restrict which entities may submit or modify meter readings reaching the Head End System. Finally, RL policy manipulation as a broader category subsumes several of the mechanisms above. Manipulation through crafted observations is addressed by the anomaly detection and adversarial training directions already discussed, manipulation through unauthorized model replacement is addressed by the Flash readout protection discussed above, and manipulation through a compromised update channel is currently precluded by the absence of an over the air update mechanism in the present prototype, as noted in “Edge deployment feasibility and hardware profiling” section. Introducing such a mechanism in future work would require cryptographically signed and verified model updates to prevent this attack vector from being reintroduced. Taken together, these considerations indicate that the current fixed policy, physically deployed architecture narrows the practical AI specific attack surface relative to a system employing online learning or remote model updates, while identifying concrete hardware and protocol level mitigations, namely Flash readout protection, IEEE 802.15.4 frame counters, and signed update verification, as priorities for hardening future field deployments.

Limitations

Several limitations of the current study should be acknowledged. First, the hardware validation was conducted at laboratory scale with short communication distances and a small number of nodes (three RF sensors, five PLC nodes). While the communication modules support significantly greater ranges, the system has not been validated at operational AMI distances or node densities. Second, the physical testbed experiments remain limited to single-run validations with a small number of physical nodes. Third, the simulated environment samples channel metrics independently within calibrated bounds, which does not fully capture the temporal correlations, impulsive noise distributions, and cyclic load-driven packet loss patterns characteristic of real PLC channels. Fourth, the current state representation is limited to three features per channel (latency, packet loss, SNR) and does not capture interference variability or temporal traffic patterns beyond the congestion model evaluated in “Simulation-based network scalability and stress testing” section. Fifth, the network-scale simulation evaluates interface selection in isolation and does not model interaction with higher-layer routing protocols, which may introduce additional coordination overhead in multi-hop scenarios. Sixth, the sensitivity analysis utilized a static, one-factor-at-a-time (OAT) parameter variation methodology. In real-world environments, joint multi-parameter changes and time-varying reward coefficient interactions may occur, which are not modeled by single-parameter sensitivity sweeps. Seventh, the communication performance evaluation is confined to the three metrics directly observable by the RL agent (latency, packet loss, SNR) plus PDR at the network scale. Extended metrics including throughput, goodput, BER, jitter, outage probability, channel utilization, retransmission overhead, and per-node energy consumption require packet-level hardware instrumentation beyond what the current testbed exposes. Eighth, the framework as evaluated assumes that PLC and RF channel observations at a given node are affected by independent random processes, and the network-scale correlated impulsive noise mechanism in “Simulation-based network scalability and stress testing” section models correlation only within the PLC medium at DCU scale. It does not explicitly model common-mode failures that could simultaneously degrade both media, such as a local transformer or DCU power outage that removes electrical continuity for PLC transmission and also interrupts power to a co-located RF gateway. In the present hierarchical architecture, both the PLC mesh and the RF star topology terminate at the same central DCU, so a DCU-level outage would interrupt connectivity for all nodes in that segment regardless of which channel the per-node policy selects; this is an architectural single point of failure that per-node interface selection cannot mitigate on its own. The Node Outage stress test in “Simulation-based network scalability and stress testing” section evaluates the case of individual nodes going fully offline and shows that aggregate network delivery is not degraded by such outages, but it does not specifically model a shared upstream event correlated across multiple nodes and both channels simultaneously. Furthermore, the simulated channel model represents degraded conditions as bounded latency, packet loss, and SNR values rather than an explicit binary unavailable state for each channel, so it cannot currently distinguish a fully unavailable medium from a severely degraded but nominally present one. Modeling DCU-level common-mode outages explicitly, together with a discrete channel-unavailable state and backup power or cellular fallback considerations, would strengthen future evaluations of this failure mode.

Future research

Future work will prioritize field trials at representative AMI distances and node densities to validate the system under real-world propagation conditions. The simulation environment will be enhanced with time-correlated channel models and coupled with network-layer routing protocols to evaluate end-to-end performance. Future comparative studies should also include representative threshold-based heuristic switching rules commonly used in hybrid PLC-RF systems. The current prototype deploys only the trained PPO actor for fixed, offline inference; deploying the full agent, including the critic and a rolling buffer of field transitions, could instead enable on-device continual learning that adapts to long-term link drift, at the cost of a larger memory footprint than the current INT8 deployment. Validation against independent hardware traces or an established network simulator such as NS-3 or OMNeT++ would also strengthen the evaluation, since the current hardware logs were retained only as calibration ranges (Table 3) rather than raw traces suitable for direct policy replay. Additional research directions include federated learning24 for privacy-preserving distributed updates, lightweight on-device training via TinyML frameworks25, hardware consolidation onto a single-chip solution, and online reward weight tuning through meta-reinforcement learning to reduce manual calibration across diverse grid topologies.