Adaptive routing protocol for large-scale power internet of things based on edge computing and federated learning

Abstract

The rapid proliferation of heterogeneous devices in Power Internet of Things (PIoT) presents significant challenges for efficient routing and energy management in large-scale deployments. This paper proposes a novel adaptive routing protocol that integrates edge computing and federated learning to address the complexities of heterogeneous device coordination and energy efficiency optimization in PIoT environments. The proposed approach employs a hierarchical architecture where edge nodes serve as distributed processing points, enabling local decision-making while maintaining global optimization through federated learning mechanisms. The adaptive routing algorithm dynamically adjusts routing parameters based on real-time network conditions, device characteristics, and energy constraints, while the federated learning framework enables collaborative optimization without centralized data sharing. Comprehensive experimental evaluation demonstrates that the proposed protocol achieves 35–50% higher network throughput, 40–60% reduction in end-to-end delay, and 45–65% energy savings compared to traditional routing protocols. The results validate the protocol’s superior scalability and robustness, maintaining consistent performance across networks of up to 10,000 devices while traditional approaches exhibit significant performance degradation beyond 1000 devices.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-41074-5.

Introduction

The rapid advancement of smart grid technologies has catalyzed the emergence of the Power Internet of Things (PIoT), which represents a paradigm shift toward intelligent, interconnected energy infrastructure¹. As global energy demands continue to escalate and renewable energy integration becomes increasingly complex, the PIoT has evolved into a critical enabler for efficient power system management and optimization². The contemporary power grid encompasses millions of heterogeneous devices, ranging from traditional power generation equipment to advanced smart meters, distributed energy resources, and intelligent sensors, creating an unprecedented level of complexity in system coordination and control³.

The heterogeneous nature of PIoT devices presents significant challenges in achieving seamless interconnection and interoperability across diverse communication protocols, data formats, and operational requirements⁴. These challenges are further compounded by the massive scale of modern power systems, where traditional centralized management approaches struggle to accommodate the dynamic and distributed characteristics of renewable energy sources, demand response systems, and electric vehicle charging infrastructure⁵. The conventional routing protocols and energy management strategies, originally designed for homogeneous network environments, demonstrate limited effectiveness when applied to the complex, multi-vendor, and multi-technology ecosystem that characterizes today’s power grid infrastructure.

Edge computing has emerged as a promising solution to address the computational and communication bottlenecks inherent in large-scale PIoT deployments⁶. By enabling distributed processing capabilities at the network edge, edge computing can significantly reduce latency, improve real-time decision-making, and enhance system resilience through decentralized intelligence distribution. The integration of edge computing with power systems offers unique opportunities to optimize local energy management, reduce bandwidth requirements for data transmission, and improve overall system reliability through localized fault detection and mitigation strategies.

Simultaneously, federated learning has gained considerable attention as a privacy-preserving machine learning paradigm that enables collaborative model training without centralized data sharing⁷. In the context of power systems, federated learning presents significant advantages for cross-organizational knowledge sharing, regulatory compliance, and distributed optimization while maintaining data sovereignty and security. The combination of federated learning with edge computing creates a powerful framework for developing intelligent, adaptive, and scalable solutions for PIoT management and optimization.

Despite these technological advances, existing research has not adequately addressed the specific challenges of developing adaptive routing protocols that can dynamically accommodate the diverse communication requirements and energy constraints of heterogeneous PIoT devices. Current approaches often assume homogeneous device characteristics or rely on static configuration parameters that fail to adapt to changing network conditions and energy availability patterns. Furthermore, the integration of edge computing and federated learning for simultaneous routing optimization and energy efficiency improvement remains largely unexplored in the power systems domain.

This paper addresses these research gaps by proposing a novel adaptive routing protocol specifically designed for large-scale PIoT environments with heterogeneous devices, complemented by an energy efficiency optimization algorithm that leverages the synergistic benefits of edge computing and federated learning⁸. The primary innovation lies in the development of a self-adaptive framework that can dynamically adjust routing decisions based on real-time device characteristics, network conditions, and energy constraints while continuously learning and improving through federated optimization processes.

The main contributions of this work include: (1) the development of a comprehensive adaptive routing protocol that accommodates the diverse communication requirements of heterogeneous PIoT devices through dynamic parameter adjustment and intelligent path selection mechanisms; (2) the design of an energy efficiency optimization algorithm that integrates edge computing capabilities with federated learning principles to achieve distributed optimization while maintaining data privacy and security; (3) the establishment of a unified framework that enables seamless coordination between routing optimization and energy management across multiple organizational boundaries; and (4) the provision of theoretical analysis and performance evaluation demonstrating the effectiveness of the proposed approach in large-scale PIoT scenarios.

The remainder of this paper is organized as follows: Section II presents a comprehensive review of related work in PIoT routing protocols, edge computing applications, and federated learning in power systems. Section III introduces the system model and problem formulation, establishing the theoretical foundation for the proposed approach. Section IV details the adaptive routing protocol design, including the dynamic parameter adjustment mechanisms and intelligent path selection algorithms. Section V describes the energy efficiency optimization algorithm, focusing on the integration of edge computing and federated learning principles. Section VI presents the experimental setup and performance evaluation results, demonstrating the effectiveness of the proposed approach through comprehensive simulation studies. Finally, Section VII concludes the paper and outlines future research directions.

Analysis of heterogeneous device characteristics in power internet of things

The Power Internet of Things encompasses a diverse ecosystem of heterogeneous devices, each exhibiting distinct communication characteristics, operational requirements, and functional capabilities that collectively contribute to the complexity of modern power grid management⁹. Smart meters, representing the most prevalent category of PIoT devices, typically employ advanced metering infrastructure (AMI) communication protocols such as ZigBee, Wi-Fi, and cellular networks, with data transmission rates ranging from kilobits per second for basic consumption readings to megabits per second for real-time power quality monitoring¹⁰. These devices generate structured data in various formats including DLMS/COSEM, IEC 61,850, and proprietary manufacturer-specific protocols, creating significant challenges for standardized data integration and processing across different vendor platforms.

Distributed sensors throughout the power grid infrastructure exhibit fundamentally different communication patterns compared to smart meters, often requiring low-latency, high-frequency data transmission for critical monitoring applications such as fault detection, load balancing, and environmental condition assessment¹¹. Temperature sensors, voltage monitors, and current transformers typically operate with sampling rates exceeding 1 kHz and utilize industrial communication protocols including Modbus, DNP3, and IEC 61,850, resulting in heterogeneous data streams that vary significantly in format, frequency, and criticality levels. The diverse communication requirements of these sensor networks create complex routing challenges, particularly when considering the need for real-time data delivery while maintaining energy efficiency across battery-powered devices.

Control devices, including programmable logic controllers (PLCs), remote terminal units (RTUs), and intelligent electronic devices (IEDs), represent another critical category of PIoT components with unique communication characteristics and operational constraints¹². These devices typically require bidirectional communication capabilities to receive control commands and transmit status information, often operating under strict latency requirements for critical control applications such as automatic generation control, demand response, and protective relay coordination. The communication protocols utilized by control devices frequently include DNP3, IEC 61,850 GOOSE messaging, and vendor-specific proprietary protocols, contributing to the overall heterogeneity of the PIoT communication landscape.

The network topology structure of PIoT systems presents additional complexity due to the hierarchical and distributed nature of power grid infrastructure, where devices may be organized in star, mesh, or hybrid configurations depending on geographical constraints, communication requirements, and reliability considerations¹³. Local area networks within substations often utilize high-speed Ethernet-based communication, while wide area networks connecting distributed resources may rely on cellular, satellite, or power line communication technologies. This multi-tier architecture creates routing challenges related to protocol translation, data aggregation, and quality of service management across different communication domains.

Interoperability challenges arise from the fundamental differences in communication protocols, data encoding schemes, and semantic interpretations across heterogeneous PIoT devices, requiring sophisticated middleware solutions and protocol translation mechanisms to enable seamless information exchange¹⁴. The lack of universal standards for data representation, communication timing, and security protocols further complicates the integration process, particularly when considering the need for real-time coordination between devices from different manufacturers and technology generations. Additionally, the varying energy constraints of battery-powered sensors, line-powered smart meters, and grid-connected control devices necessitate adaptive routing strategies that can dynamically balance communication requirements with energy efficiency objectives, creating multi-objective optimization problems that traditional routing protocols are ill-equipped to address effectively.

Theoretical foundation of edge computing and federated learning

Edge computing represents a distributed computing paradigm that brings computational resources closer to data sources, enabling real-time processing and reducing latency through strategically positioned edge nodes throughout the network infrastructure¹⁵. The hierarchical architecture of edge computing systems typically consists of three primary layers: the device layer containing IoT sensors and actuators, the edge layer encompassing local processing nodes and gateways, and the cloud layer providing centralized coordination and long-term storage capabilities. The computational resource allocation strategy in edge computing environments can be mathematically formulated as an optimization problem that seeks to minimize the total system cost while satisfying service quality requirements:

where represents the communication cost between device and edge node , is a binary variable indicating task assignment, denotes the processing cost at edge node , and represents the activation status of edge node ¹⁶.

The dynamic resource allocation mechanism in edge computing systems must consider both computational capacity constraints and communication bandwidth limitations, leading to a multi-objective optimization framework that balances processing efficiency with energy consumption¹⁷. The resource allocation decision can be expressed through the following constraint optimization model:

where represents the computational resource requirement of task , and denotes the computational capacity of edge node .

Federated learning fundamentally transforms the traditional centralized machine learning paradigm by enabling collaborative model training across distributed devices without requiring centralized data sharing, thereby addressing critical privacy and security concerns in power grid applications¹⁸. The federated learning algorithm operates through iterative local model training followed by global model aggregation, where each participating device trains a local model using its private dataset and subsequently shares only the model parameters with the central aggregator. The global model update mechanism can be mathematically represented as:

where represents the global model parameters at iteration , denotes the local model parameters from device , is the number of data samples at device , and represents the total number of samples across all participating devices.

The privacy protection mechanism in federated learning systems employs differential privacy techniques and secure aggregation protocols to ensure that individual device data remains confidential while enabling effective collaborative learning¹⁹. The differential privacy mechanism adds calibrated noise to the model updates, which can be formulated as:

where represents the noisy model parameters, denotes Gaussian noise with variance , is the sensitivity parameter, and is the identity matrix.

The integration of edge computing and federated learning creates a synergistic framework that addresses the computational limitations of individual IoT devices while preserving data privacy and reducing communication overhead²⁰. Edge nodes serve as intermediate aggregation points for federated learning, enabling hierarchical model training that reduces the communication burden on central servers while maintaining learning effectiveness. This combined approach offers particular advantages in power grid applications where real-time decision-making, data privacy, and distributed intelligence are critical requirements. The edge-federated learning architecture enables local model training at substations and distribution points, regional model aggregation at edge nodes, and global optimization at the system level, creating a multi-tier learning framework that can adapt to varying computational resources and communication constraints across different levels of the power grid hierarchy.

Adaptive routing protocols and energy efficiency optimization methods

Traditional routing protocols, originally designed for homogeneous network environments, exhibit significant limitations when applied to large-scale Power Internet of Things deployments due to their static nature and inability to accommodate the diverse communication requirements of heterogeneous devices²¹. Conventional protocols such as AODV (Ad-hoc On-Demand Distance Vector) and DSR (Dynamic Source Routing) rely on predetermined routing metrics that fail to consider the unique characteristics of power grid devices, including varying energy constraints, communication priorities, and real-time requirements. These protocols typically employ shortest-path algorithms that optimize for minimal hop count or distance, neglecting critical factors such as device battery levels, communication channel quality, and dynamic network conditions that are prevalent in PIoT environments.

Adaptive routing algorithms represent a paradigm shift toward intelligent, context-aware routing decisions that can dynamically adjust to changing network conditions and device characteristics²². These algorithms incorporate multiple routing metrics including residual energy, link quality, traffic load, and device capabilities to make optimal routing decisions in real-time. The adaptive routing process typically employs machine learning techniques such as reinforcement learning or genetic algorithms to continuously optimize routing tables based on network performance feedback and historical data patterns. The multi-metric routing decision framework enables the protocol to balance conflicting objectives such as energy efficiency, latency minimization, and reliability enhancement through sophisticated optimization techniques.

Dynamic topology management technology addresses the challenges associated with frequent network changes in PIoT environments, where devices may join or leave the network, experience varying communication conditions, or undergo maintenance cycles that affect network connectivity²³. Advanced topology management systems employ predictive analytics and machine learning algorithms to anticipate network changes and proactively adjust routing strategies to maintain optimal performance. These systems typically implement distributed consensus mechanisms that enable network nodes to collaboratively maintain consistent topology information without requiring centralized coordination, thereby improving system resilience and scalability.

Energy efficiency optimization strategies in PIoT networks focus on minimizing overall system power consumption while maintaining acceptable service quality levels across all connected devices²⁴. These strategies encompass multiple optimization dimensions including transmission power control, duty cycle management, data aggregation techniques, and sleep scheduling algorithms that coordinate device operations to reduce unnecessary energy expenditure. The optimization process typically considers the heterogeneous energy profiles of different device types, ranging from battery-powered sensors with severe energy constraints to line-powered smart meters with abundant energy resources.

Power consumption modeling methods provide the theoretical foundation for quantitative energy efficiency analysis and optimization in PIoT systems²⁵. These models typically decompose total device power consumption into distinct components including communication power, processing power, and idle power, enabling precise prediction of energy usage patterns under different operational scenarios. The modeling approach often employs statistical analysis of historical power consumption data combined with physics-based models that account for factors such as transmission distance, data payload size, and environmental conditions. Advanced power models incorporate stochastic elements to capture the variability in energy consumption patterns across different device types and operational conditions.

The integration of adaptive routing protocols with energy efficiency optimization creates a comprehensive framework for intelligent network management that can simultaneously address the challenges of heterogeneous device coordination, dynamic network conditions, and energy constraints²⁶. This integrated approach enables the development of multi-objective optimization algorithms that can balance competing requirements such as communication reliability, energy efficiency, and system performance through sophisticated mathematical modeling and machine learning techniques. The theoretical foundation established through these methods provides the necessary analytical tools for designing and evaluating advanced routing protocols specifically tailored to the unique requirements of large-scale PIoT deployments, enabling the development of robust, scalable, and energy-efficient communication solutions for modern power grid infrastructure.

System architecture and protocol framework design

The proposed adaptive routing protocol leverages a hierarchical edge computing architecture specifically designed to address the scalability and heterogeneity challenges inherent in large-scale Power Internet of Things deployments²⁷. The system architecture consists of four distinct layers: the device layer containing heterogeneous PIoT devices, the edge layer comprising distributed edge computing nodes, the fog layer providing regional coordination and aggregation services, and the cloud layer enabling centralized management and long-term analytics. This multi-tier architecture enables distributed intelligence deployment while maintaining efficient coordination across different abstraction levels, thereby reducing communication latency and improving system responsiveness for critical power grid operations.

The comprehensive system architecture, as presented in Fig. 1, illustrates the integration of edge computing nodes with federated learning mechanisms to create a robust communication framework for heterogeneous device coordination. The architecture demonstrates how edge nodes serve as intermediate processing and aggregation points, enabling local decision-making while maintaining global system coherence through federated learning protocols. Each edge node maintains localized routing tables and device profiles, enabling rapid route calculation and traffic forwarding without requiring centralized coordination for routine operations.

Fig. 1.

Edge computing and federated learning-based power internet of things system architecture.

The federated learning-driven adaptive routing protocol framework incorporates distributed machine learning capabilities to continuously optimize routing decisions based on real-time network conditions and historical performance data²⁸. The framework employs a hierarchical federated learning structure where edge nodes serve as local aggregators for device-level model training, while fog nodes coordinate regional model updates and cloud servers maintain global model consistency. This distributed learning approach enables the protocol to adapt to changing network conditions, device characteristics, and traffic patterns without requiring centralized data collection or processing.

The diverse characteristics of PIoT devices necessitate a comprehensive understanding of their communication parameters and operational requirements, as demonstrated in Table 1, which provides a detailed comparison of device types, communication protocols, data rates, power consumption levels, and priority classifications. The table reveals significant variations in device capabilities, ranging from low-power sensors with minimal data transmission requirements to high-performance control devices requiring real-time communication capabilities. These heterogeneous characteristics directly influence the routing protocol design, requiring adaptive algorithms that can dynamically adjust routing decisions based on device-specific constraints and requirements.

Table 1.

Device type and communication parameter comparison.

Device type	Communication protocol	Data rate	Power level	Priority	Latency requirement
Smart meter	ZigBee/Wi-Fi	1–10 Mbps	Medium	Normal	< 1000ms
Temperature sensor	LoRaWAN	0.3–50 kbps	Low	Low	< 5000ms
Voltage monitor	DNP3/IEC61850	1–100 kbps	High	High	< 100ms
Current transformer	IEC61850 GOOSE	10–100 Mbps	High	Critical	< 10ms
PLC controller	Modbus/DNP3	1–10 Mbps	High	Critical	< 50ms
RTU Device	DNP3/IEC61850	100 kbps-1 Mbps	Medium	High	< 200ms
Environmental sensor	IEEE 802.15.4	20–250 kbps	Low	Low	< 2000ms
Protection relay	IEC61850	1-100 Mbps	High	Critical	< 5ms

The device discovery mechanism implements a multi-phase discovery protocol that combines passive listening, active probing, and collaborative information sharing to maintain accurate device inventories across the distributed network²⁹. The discovery process begins with passive monitoring of network traffic to identify active devices and their communication patterns, followed by targeted probing to gather detailed device capabilities and requirements. The collaborative information sharing phase leverages the federated learning framework to distribute device discovery information across edge nodes, enabling rapid network adaptation when new devices join or existing devices modify their operational parameters.

The path selection algorithm employs a multi-objective optimization approach that considers multiple routing metrics including energy efficiency, communication latency, reliability, and device capabilities³⁰. The algorithm can be mathematically formulated as a constrained optimization problem:

where represents the energy cost of routing from device to device , denotes the communication latency, indicates the link reliability, represents the device capability compatibility factor, and are weighting coefficients that balance the different optimization objectives.

The load balancing strategy implements a distributed approach that considers both traffic distribution and energy consumption patterns across the network³¹. The strategy employs a dynamic load assessment mechanism that can be expressed as:

where represents the current traffic load at node , denotes the current energy consumption rate, indicates the queue length, and are balancing parameters that weight the different load components.

The protocol framework incorporates an adaptive threshold management system that dynamically adjusts routing parameters based on network conditions and device performance metrics. This system employs machine learning algorithms to predict optimal threshold values and routing parameters based on historical data and real-time network state information. The adaptive threshold mechanism can be mathematically represented as:

where represents the current threshold parameters, is the learning rate, and denotes the gradient of the objective function with respect to the threshold parameters.

To bridge the gap between node-state prediction and concrete routing decisions, we now describe exactly how predicted network states translate into routing actions. After the Bi-LSTM model outputs the predicted network state vector , the routing module extracts four key components: predicted traffic volume , predicted energy level , predicted link quality , and predicted queue length . These predictions are converted into routing weight coefficients through the following mapping functions: the energy weight is adjusted as when predicted energy falls below the threshold; the latency weight is scaled by when high traffic is anticipated; and the reliability weight is modified as when link quality degradation is predicted. Here, , , and are sensitivity parameters empirically set to 0.5, 0.3, and 0.4 respectively. These dynamically adjusted weights feed directly into the path selection algorithm the path selection algorithm, determining which candidate path is chosen.

Beyond per-node predictions, the attention mechanism in our Bi-LSTM architecture captures multi-node interactions and neighborhood-level patterns. The 8-head attention layer computes correlation scores between nodes, effectively learning which neighboring nodes influence each other’s traffic patterns. During path selection, the routing algorithm queries the attention weights to identify potential contention: if two candidate paths share nodes with high mutual attention scores and both are predicted to experience elevated traffic, the algorithm preferentially selects a path with lower predicted contention to exploit path diversity. Furthermore, topology-level information is incorporated through a graph-based neighbor aggregation step, where each node’s predicted state is refined by averaging predictions from its one-hop neighbors weighted by link reliability, ensuring that routing decisions account for neighborhood structure rather than treating nodes in isolation.

The integration of edge computing capabilities with federated learning principles enables the protocol to achieve distributed optimization while maintaining data privacy and security across different organizational boundaries within the power grid infrastructure. The system components interact through well-defined mechanisms: the device discovery module feeds discovered device capabilities into the routing algorithm’s feature vector for capability-aware path selection; edge nodes serve dual roles as federated learning clients for training local routing models and as regional aggregators coordinating model updates; the network state predictor generates traffic forecasts that influence real-time routing decisions through dynamic weight adjustment; the energy optimizer coordinates with the routing protocol to assign lower priority to energy-constrained nodes and trigger sleep scheduling for underutilized devices; and federated aggregation occurs at the fog layer every 30 s before broadcasting updated models network-wide to synchronize routing tables. The framework supports dynamic protocol adaptation based on changing network conditions, device characteristics, and operational requirements, thereby ensuring robust and efficient communication in large-scale PIoT environments. The distributed nature of the protocol reduces dependency on centralized coordination, improving system resilience and scalability while enabling real-time decision-making capabilities essential for critical power grid applications.

Federated learning-driven dynamic routing optimization algorithm

The proposed federated learning-driven dynamic routing optimization algorithm marks a departure from traditional centralized routing approaches, introducing distributed decision-making capabilities that harness the collective intelligence of heterogeneous PIoT devices³². Rather than relying on a single central controller, this algorithm builds on a collaborative learning framework in which edge nodes periodically train local routing models. To be precise, each edge node maintains a fixed-size sliding window buffer containing the most recent 10,000 time-series samples of observed network conditions, traffic patterns, and device performance metrics. When new observations arrive, older samples are discarded in a first-in-first-out manner, ensuring that the training dataset reflects current network dynamics while remaining bounded in size. Every 30 s, each edge node performs local training on this fixed buffer for exactly 5 epochs, where one epoch constitutes a complete pass over all samples currently stored in the buffer. Model parameters are then shared through secure aggregation protocols to achieve global routing optimization without compromising data privacy. This periodic training approach, as opposed to truly online gradient updates, allows the system to adapt to evolving network conditions while preserving the confidentiality of sensitive power grid operational data and maintaining computational tractability on resource-constrained edge devices.

The comprehensive algorithm workflow, as illustrated in Fig. 2, demonstrates the iterative process of local model training, parameter aggregation, and global model updates that collectively optimize routing decisions across the entire PIoT network. The flowchart reveals the multi-stage optimization process, beginning with local data collection and preprocessing at individual edge nodes, followed by local model training using historical routing performance data, secure parameter sharing through federated aggregation, and finally global model synchronization that ensures consistent routing policies across all network nodes. This structured approach enables continuous learning and adaptation while maintaining system stability and performance consistency.

Fig. 2.

Federated learning-driven dynamic routing optimization algorithm flowchart.

The network state prediction model employs a bidirectional Long Short-Term Memory (Bi-LSTM) network with multi-head attention mechanism to capture temporal dependencies in network traffic patterns and device behavior³³. The model architecture consists of five layers: an input layer accepting time-series feature vectors of dimension 128 (including normalized traffic volume, device energy percentages, link quality indicators measured by RSSI and SNR, and normalized queue lengths); a two-layer Bi-LSTM with 256 hidden units per direction enabling bidirectional temporal dependency capture; a multi-head attention layer with 8 heads of dimension 64 each computing attention scores as Attention(Q, K,V) = softmax(QK^T/√d_k)V where Q, K, V are query, key, and value matrices; two fully connected layers with 512 and 256 neurons respectively using ReLU activation and 0.3 dropout rate; and an output layer generating predicted network state vectors for the next time step with the same dimensionality as input features. The model is implemented using PyTorch 1.12 and integrated into NS-3 through Python bindings for real-time prediction during simulation. The prediction model processes multiple input features including historical traffic volumes spanning 50 time steps at 0.1-second intervals, device energy levels, link quality indicators, and environmental conditions to forecast future network states with high accuracy. The model architecture enables the system to anticipate network congestion, device failures, and communication bottlenecks before they occur, thereby enabling proactive routing adjustments that maintain optimal network performance.

The mathematical formulation of the network state prediction model can be expressed as:

where represents the predicted network state at time , denotes the current network state, indicates the current traffic matrix, represents the energy status vector, denotes the link quality matrix, and represents the learned model parameters through federated learning. Input features undergo comprehensive preprocessing: all numerical features are normalized using min-max scaling to the [0,1] range; categorical features including device types and protocol types are one-hot encoded; multi-modal inputs are concatenated into unified feature vectors of dimension 128; and temporal data uses sliding windows of 50 time steps with 0.1-second intervals. Labels are defined as next-step network state predictions (continuous values) for the primary task, binary congestion indicators (threshold at 80% link utilization) for auxiliary classification, and packet delivery ratios as regression targets for quality assessment. The data exhibits natural non-IID distribution across the network as each edge node observes different local traffic patterns, with 50 edge nodes each acting as a federated learning client holding 5,000–15,000 time-series sequences (average 10,000 sequences per client). Each client employs a temporal 70%/15%/15% train/validation/test split, and the system uses synchronous federated learning with 100% client participation per 30-second aggregation round over 100 total global training iterations.

The adaptive weight update mechanism implements a sophisticated parameter adjustment strategy that considers both local network conditions and global optimization objectives³⁴. The mechanism employs a multi-objective optimization approach that balances routing efficiency, energy consumption, and network reliability through dynamic weight adjustment based on real-time performance feedback. The weight update process incorporates momentum-based optimization and adaptive learning rates to ensure rapid convergence while maintaining system stability during network transitions.

The weight update mechanism can be mathematically represented as:

where represents the current weight parameters, denotes the adaptive learning rate, indicates the gradient of the loss function, and represents the momentum coefficient that enables smooth convergence.

The algorithm performance comparison based on NS-3 simulations with 1000-node network topology over 30 independent runs, as demonstrated in Table 2, reveals significant improvements in key routing metrics when compared to traditional routing protocols, particularly in terms of convergence time, optimal path length, and computational efficiency. The table shows that the proposed federated learning-driven approach achieves faster convergence times compared to centralized algorithms while maintaining comparable or superior path optimization results. The computational complexity refers to per-iteration routing decision complexity after model convergence rather than model training FLOPs, where FL-DRO achieves O(N log N) by using tree-based data structures for route lookup while the one-time training cost of O(N²·M) for M training iterations is amortized over the operational lifetime, whereas traditional algorithms require O(N²) or O(N³) computation for every routing decision. The computational complexity analysis indicates that the distributed nature of the algorithm reduces the overall system computational burden while improving scalability for large-scale PIoT deployments.

Table 2.

Routing algorithm performance parameter comparison*.

Algorithm name	Convergence time (s)	Average path length	Computational complexity†	Network overhead
Traditional AODV	15.2 ± 1.8	4.8 ± 0.3 hops	O(N²)	High
DSR Protocol	12.7 ± 1.5	4.5 ± 0.4 hops	O(N³)	Medium
OLSR Protocol	18.9 ± 2.1	5.2 ± 0.5 hops	O(N²)	High
Centralized ML	8.3 ± 0.9	3.9 ± 0.2 hops	O(N³)	Very high
Proposed FL-DRO	6.8 ± 0.7	3.7 ± 0.2 hops	O(N log N)	Low
Hybrid Approach	9.5 ± 1.1	4.1 ± 0.3 hops	O(N²)	Medium

*Results based on NS-3 simulations with 1000-node network topology, averaged over 30 independent runs. Traditional protocols (AODV, DSR, OLSR) involve communication-only energy. Centralized ML includes raw data transmission overhead (2 MB per node per round) plus central server training energy (150 W server power, amortized across devices). FL-DRO includes local training energy (2 W edge node power, 5 epochs × 12s per epoch per round) plus model parameter exchange (2.4 MB per round).

†Computational complexity refers to per-iteration routing decision complexity after model convergence, not training FLOPs. FL-DRO achieves O(N log N) through tree-based route lookup structures; one-time training cost O(N²·M) for M iterations is amortized over operational lifetime.

The real-time optimization component of the algorithm implements a continuous monitoring and adjustment mechanism that responds to network condition changes within milliseconds³⁵. The system employs a distributed event-driven architecture where edge nodes continuously monitor local network conditions and trigger optimization processes when predefined thresholds are exceeded. This proactive approach enables the algorithm to prevent network congestion and maintain optimal routing performance even during high-traffic periods or device failures.

The congestion avoidance strategy integrates predictive analytics with real-time traffic management to prevent network bottlenecks before they impact system performance³⁶. The strategy employs a multi-path routing approach that distributes traffic across multiple routes based on predicted congestion levels and available network capacity. The congestion avoidance mechanism can be formulated as:

where represents the probability of selecting path from node , denotes the congestion level on path , indicates the reliability of path , and are tuning parameters that balance congestion avoidance with reliability requirements.

The federated learning framework ensures that routing optimization knowledge is continuously shared across the network while maintaining data privacy and security through differential privacy mechanisms and secure aggregation protocols. To address the inherent non-IID data distribution in Power IoT environments where different edge nodes observe heterogeneous traffic patterns, we employ the FedProx algorithm instead of standard FedAvg by adding a proximal term to the local objective function: with µ = 0.01 to control divergence from the global model³⁷. Additionally, we implement performance-based adaptive aggregation weights calculated as rather than simple sample-size-based weighting, limit local training to 5 epochs per round to prevent overfitting to local distributions, apply data augmentation through temporal jittering and noise injection to increase local data diversity, and monitor per-client validation loss to implement early stopping when global model accuracy plateaus. The distributed nature of the algorithm enables rapid adaptation to changing network conditions while reducing the communication overhead associated with centralized optimization approaches. The integration of predictive analytics with real-time optimization enables the system to maintain optimal routing performance even under dynamic network conditions, thereby improving overall network transmission efficiency and reliability in large-scale PIoT environments.

Energy efficiency optimization and power management strategy

The power-aware scheduling model based on edge computing addresses the critical challenge of minimizing energy consumption while maintaining communication quality in large-scale PIoT deployments through intelligent task allocation and resource management strategies³⁸. The model incorporates comprehensive energy profiling of heterogeneous devices, considering both static power consumption from idle operations and dynamic power consumption from active communication and processing tasks. The scheduling framework employs a multi-objective optimization approach that balances energy efficiency with service quality requirements, enabling the system to adapt power consumption patterns based on real-time network conditions and application priorities.

The dynamic power allocation algorithm implements a sophisticated resource management strategy that optimizes power distribution across edge nodes and connected devices based on current workload demands and energy availability³⁹. The algorithm continuously monitors device energy levels, communication requirements, and computational loads to make intelligent power allocation decisions that minimize overall system energy consumption while preventing device failures due to energy depletion. The power allocation process can be mathematically formulated as:

where represents the total system power consumption, denotes the communication power consumption of device , indicates the computational power consumption, and represents the idle power consumption.

The sleep-wake mechanism introduces intelligent duty cycling strategies that coordinate device operations to minimize unnecessary power consumption during low-activity periods while ensuring rapid response to critical events⁴⁰. The mechanism employs predictive analytics to forecast communication patterns and device usage requirements, enabling proactive scheduling of sleep and wake cycles that optimize energy consumption without compromising system responsiveness. The sleep-wake scheduling algorithm considers multiple factors including device energy levels, communication priorities, and network topology constraints to determine optimal sleep durations and wake-up schedules for each device.

The load migration strategy implements dynamic workload redistribution mechanisms that transfer computational and communication tasks from energy-constrained devices to nodes with abundant power resources⁴¹. The strategy employs real-time monitoring of device energy levels and performance metrics to identify optimal migration opportunities and execute seamless task transfers that maintain service continuity while reducing energy consumption. The load migration process considers factors such as migration overhead, communication latency, and device capabilities to ensure that the energy savings from migration exceed the costs associated with task transfer.

The comprehensive comparison of energy efficiency optimization strategies based on simulation results averaged over 30 independent runs with 5000-second duration each, as presented in Table 3, demonstrates the trade-offs between different approaches in terms of energy savings, response time, implementation complexity, and resource overhead. Response time includes both routing computation time and model inference time (forward pass for ML-based methods), while energy values represent average per-packet transmission energy plus computational energy measured using NS-3’s extended Energy Framework module. The table reveals that advanced optimization strategies such as the proposed federated learning-based approach achieve superior energy savings compared to traditional methods while maintaining acceptable response times and moderate implementation complexity. The analysis indicates that intelligent scheduling and dynamic power allocation strategies provide the most significant energy efficiency improvements, particularly in scenarios with heterogeneous device types and varying workload patterns.

Table 3.

Energy efficiency optimization strategy comparison*.

Optimization strategy	Energy saving rate (%)†	Response time (ms)‡	Applicable scenarios	Implementation complexity	Resource overhead
Static scheduling	18.3 ± 2.7	72.4 ± 8.1	Homogeneous networks	Low	Low
Dynamic load balancing	28.7 ± 3.2	51.3 ± 6.8	Medium-scale PIoT	Medium	Medium
Adaptive duty cycling	36.2 ± 4.1	38.7 ± 5.2	Battery-powered Devices	Medium	Low
Predictive migration	46.8 ± 3.9	27.6 ± 4.3	Heterogeneous Networks	High	High
Federated learning-based	56.4 ± 4.5	21.2 ± 3.1	Large-scale PIoT	High	Medium

*Simulation results averaged over 30 independent runs, each with 5000-second duration. All strategies use identical communication energy models (WiFi Radio Energy Model). Static Scheduling and Dynamic Load Balancing involve no ML computation. Adaptive Duty Cycling uses rule-based logic with negligible computational overhead. Predictive Migration uses lightweight linear regression models. Federated Learning-based uses Bi-LSTM with edge node power consumption modeled at 2 W based on ARM Cortex-A53 class embedded devices^42–44.

†Energy saving rate compared to baseline routing without optimization. Values shown as mean ± standard deviation.

‡Response time includes routing computation and model inference (forward pass). ML training costs amortized over operational lifetime.

The optimization framework integrates multiple energy management techniques including transmission power control, computational resource scaling, and intelligent caching strategies to achieve comprehensive energy efficiency improvements⁴⁵. The framework employs machine learning algorithms to continuously learn and adapt energy consumption patterns based on historical data and real-time network conditions. The overall energy optimization objective can be expressed as:

subject to quality of service constraints and device operational requirements, where represents the energy cost coefficient for device , denotes the energy consumption of device at time , and represents the optimization time horizon.

The power management strategy implements hierarchical control mechanisms that coordinate energy optimization across multiple system layers, from individual device power management to network-wide energy coordination. The strategy employs distributed consensus algorithms to ensure consistent power management policies across all network nodes while maintaining local autonomy for real-time power allocation decisions. The integration of edge computing capabilities enables localized power optimization that reduces communication overhead and improves system responsiveness, while federated learning principles ensure continuous improvement of energy management strategies through collaborative learning across the network. The comprehensive approach achieves significant energy efficiency improvements while maintaining high communication quality and system reliability, making it particularly suitable for large-scale PIoT deployments where energy constraints are critical operational considerations.

Simulation environment setup and parameter configuration

The large-scale Power Internet of Things simulation platform was constructed using Network Simulator 3 (NS-3) version 3.35 framework enhanced with custom modules to accurately model the heterogeneous characteristics and communication protocols of PIoT devices⁴⁶. NS-3 does not natively support federated learning, so we extended it with custom C + + and Python modules (approximately 5,000 lines of code) including PyTorch 1.12 integration via Python bindings for machine learning model implementation, custom application-layer protocols for federated aggregation, extended energy models for computational cost tracking, and synchronization protocols for model parameter exchange between edge nodes. The simulation environment incorporates realistic power grid topologies based on IEEE standard test systems, including distribution networks, transmission substations, and renewable energy integration points to provide comprehensive evaluation scenarios. The platform supports simultaneous simulation of up to 10,000 heterogeneous devices distributed across multiple network clusters, enabling scalability assessment of the proposed adaptive routing protocol under various network conditions and device densities.

The network topology configuration employs a hierarchical three-level tree structure with 1 central cloud coordination server at level 1, 10 fog nodes at level 2 (each serving 5 edge nodes), and 50 edge nodes at level 3 (each managing 20 IoT devices) for a baseline of 1000 total devices expandable to 10,000 in scalability tests, with uniform distribution within a 1 km² simulation area. The topology encompasses multiple operational domains within the power grid infrastructure, including generation facilities, transmission networks, distribution systems, and customer premises equipment. Each network cluster contains a mixture of device types with varying communication requirements, energy constraints, and operational priorities to accurately represent real-world PIoT deployments. The simulation incorporates realistic channel models including log-distance path loss with exponent 3.5, log-normal shadowing with σ = 4dB standard deviation, Rayleigh fading for multipath effects, and SINR-based reception with − 90dBm thermal noise floor that account for path loss, fading, and interference effects commonly encountered in power grid communication environments.

The comprehensive experimental parameter configuration, as detailed in Table 4, establishes the simulation parameters that govern device behavior, network characteristics, and protocol performance evaluation. The table provides the complete specification of simulation parameters including device quantities, communication ranges, data rates, energy levels, federated learning hyperparameters (FedProx with µ = 0.01, learning rate 0.01 with cosine annealing, 5 local epochs per round, batch size 32, 100 global rounds, 100% client participation), RNN model parameters (Bi-LSTM hidden dimension 256, 2 LSTM layers, 8 attention heads, 0.3 dropout, 50-step sequences at 0.1s intervals), and routing protocol-specific settings (1s route update interval, 8-hop maximum, weight coefficients α₁=0.3, α₂=0.25, α₃=0.25, α₄=0.2, 0.8 load balancing threshold) that ensure reproducible and comprehensive evaluation of the proposed adaptive routing protocol. The parameter ranges are selected to cover typical operational scenarios encountered in real-world PIoT deployments while enabling systematic performance analysis across different network scales and conditions. Each experiment is repeated 30 times with different random seeds to ensure statistical reliability, with results reported as mean values with 95% confidence intervals.

Table 4.

Experimental parameter configuration*.

Parameter name	Value range	Default value
Network size (devices)	100–10000	1000
Communication range (m)	50–500	200
Data rate (Mbps)	0.1–100	10
Initial energy (J)	1000–50,000	10,000
Simulation time (s)	1000–10,000	5000
Edge node count	10–100	50
Packet size (bytes)	64–1500	512
Traffic load (packets/s)	1–100	10
Propagation model	–	Log-distance (n = 3.5)
Shadowing model	–	Log-normal (σ = 4dB)
Fading model	–	Rayleigh
Thermal noise (dBm)	–	−90
FL Algorithm	–	FedProx
Proximal Term µ	0.001-0.1	0.01
Learning rate	0.001-0.1	0.01
Learning rate decay	–	Cosine annealing
Local Epochs	1–10	5
Batch size	16–64	32
Global rounds	50–200	100
Client participation	50–100%	100%
Aggregation period (s)	10–100	30
RNN type	–	Bi-LSTM
LSTM hidden dimension	128–512	256
LSTM layers	1–3	2
Attention heads	4–16	8
Dropout rate	0.1–0.5	0.3
Sequence length	20–100	50
Time step (s)	0.05–0.5	0.1
Route update interval (s)	0.5-5	1
Maximum hop count	5–15	8
Weight α₁ (energy)	0.2–0.4	0.3
Weight α₂ (latency)	0.2–0.4	0.25
Weight α₃ (reliability)	0.2–0.4	0.25
Weight α₄ (capability)	0.1–0.3	0.2
Load balance threshold	0.6–0.9	0.8
Number of runs	–	30

*All experiments repeated 30 times with different random seeds. Results reported with 95% confidence intervals.

The performance evaluation framework defines multiple key metrics including network throughput, end-to-end delay, packet delivery ratio, energy consumption, and routing convergence time to provide comprehensive assessment of protocol effectiveness. Energy consumption is measured using NS-3’s Energy Framework module extended with custom energy models. Communication energy uses the WiFi Radio Energy Model with transmission power 0.5 W, reception power 0.3 W, idle power 0.1 W, and sleep power 0.01 W. Computational energy for machine learning operations requires careful modeling that accounts for the embedded nature of edge devices in PIoT deployments.

We emphasize that our edge nodes are modeled as low-power embedded devices based on ARM Cortex-A53 class processors, not desktop x86 CPUs or GPU servers. Prior studies have demonstrated that CNN training on high-performance x86 CPUs can exceed 100 W⁴², but such measurements do not apply to our embedded edge computing scenario. For embedded edge devices with ARM processors, power consumption during neural network operations typically ranges from 1 to 5 W depending on workload intensity⁴³. The NXP i.MX 8 M+ edge processor, which integrates a Cortex-A53 CPU with an NPU, consumes approximately 2 W while delivering 2.3 TOPS of neural network performance⁴⁴. Similarly, NVIDIA Jetson Nano devices consume 5–10 W for full inference workloads, while lighter ARM-based systems operate in the 1–3 W range⁴³. Based on these references and considering our relatively lightweight Bi-LSTM model with 256 hidden units, we adopt as a representative power consumption value for edge node processing during training operations.

Computational energy is modeled as , where represents edge node processing power, seconds denotes the empirically measured time for one complete training epoch over the 10,000-sample buffer on our modeled Cortex-A53 edge device, and is the number of local epochs per federated learning round. Inference energy is modeled as where per prediction. The total training energy per round per edge node is therefore , which occurs every 30 s.

For baseline algorithm energy modeling, we apply consistent methodology across all compared methods. Traditional routing protocols (AODV, DSR, OLSR) involve only communication energy since they perform no machine learning computations; their energy consumption is calculated as using identical WiFi Radio Energy Model parameters. The Centralized ML baseline incurs both communication and computation costs: edge devices transmit raw training data to a central server (modeled as additional communication overhead of 2 MB per device per training round), and the central server performs training with power consumption scaled by network size. The central server is modeled as a high-performance computing node with based on typical server CPU power consumption during ML training⁴², and training time scales linearly with aggregated dataset size. Total centralized ML energy is computed as , where the server training cost is amortized across all participating devices. This modeling explains why the centralized approach exhibits higher per-device energy consumption despite potentially superior model quality, as the communication overhead for raw data transmission dominates energy costs in large-scale deployments.

Total system energy is computed as with ML training costs amortized over 100 global rounds. The evaluation accounts for both communication and computational energy. To validate our energy modeling assumptions, we conducted sensitivity analysis varying from 1 W to 5 W, finding that our conclusions regarding relative protocol performance remain robust across this range, with absolute energy values scaling proportionally. Federated learning overhead represents approximately 15% of total system energy consumption, which is offset by the 45–65% energy savings from intelligent routing optimization that reduces unnecessary packet retransmissions and avoids congested paths. The evaluation methodology compares the proposed federated learning-driven adaptive routing protocol against established baseline algorithms including AODV, DSR, OLSR, and centralized machine learning approaches to demonstrate performance improvements and validate the effectiveness of the edge computing and federated learning integration.

The protocol performance comparison across different network scales, as presented in Fig. 3, demonstrates the scalability advantages of the proposed approach compared to traditional routing protocols. The figure shows six performance metrics (throughput, delay, packet loss, energy consumption, convergence time, and scalability score) for five algorithms across network sizes from 100 to 10,000 devices, with each algorithm clearly distinguished by unique line styles and colors: red solid line for proposed FL-DRO, blue dashed line for AODV, green dotted line for DSR, orange dash-dot line for OLSR, and purple short-dash line for Centralized ML. The figure reveals that the proposed protocol maintains superior performance metrics including lower latency, higher throughput, and reduced energy consumption as network size increases, while baseline protocols show significant performance degradation with increased network scale. Error bars represent 95% confidence intervals computed from 30 independent simulation runs. The comparison clearly illustrates the benefits of distributed intelligence and adaptive optimization in large-scale PIoT environments.

Fig. 3.

Protocol performance comparison across different network scales. All subplots include error bars representing 95% confidence intervals from 30 independent runs. Red solid = Proposed FL-DRO, Blue dashed = AODV, Green dotted = DSR, Orange dash-dot = OLSR, Purple short-dash = Centralized ML. Energy consumption (subplot d) includes both communication energy (identical WiFi Radio Energy Model across all protocols) and computational energy where applicable. For Centralized ML, energy includes raw data upload costs and amortized server training energy. For FL-DRO, energy includes local training on 2 W edge devices and model parameter synchronization. The superior energy performance of FL-DRO despite ML overhead reflects reduced retransmissions and optimized path selection enabled by intelligent routing.

The baseline algorithms include traditional reactive and proactive routing protocols as well as recent machine learning-based approaches to ensure comprehensive performance evaluation⁴⁷. The experimental methodology employs statistical analysis with confidence intervals and multiple independent simulation runs to ensure result reliability and reproducibility. The simulation framework incorporates realistic device failure models, dynamic topology changes, and varying traffic patterns to evaluate protocol robustness under challenging operational conditions commonly encountered in power grid environments.

Routing protocol performance evaluation

The comprehensive performance evaluation demonstrates significant improvements of the proposed federated learning-driven adaptive routing protocol compared to traditional routing methods across multiple critical performance metrics including network throughput, end-to-end delay, and packet loss rate⁴⁸. The experimental results reveal that the adaptive mechanism enables dynamic optimization of routing decisions based on real-time network conditions, resulting in substantial performance enhancements particularly in large-scale heterogeneous PIoT environments where traditional protocols exhibit significant performance degradation.

The network throughput analysis indicates that the proposed protocol achieves 35–50% higher throughput compared to conventional routing protocols such as AODV and DSR, primarily due to the intelligent path selection mechanism that considers multiple routing metrics simultaneously. The adaptive routing algorithm effectively distributes network traffic across multiple paths while avoiding congested nodes, resulting in improved overall network utilization and reduced bottleneck formation. The federated learning component enables continuous optimization of routing decisions based on historical performance data, leading to progressively improved throughput performance as the system learns and adapts to network patterns.

The end-to-end delay evaluation demonstrates significant latency reduction achieved through the edge computing-based distributed processing architecture and predictive routing optimization. The proposed protocol reduces average packet delay by 40–60% compared to traditional centralized approaches, primarily through localized decision-making at edge nodes and proactive route optimization based on predicted network conditions. The distributed nature of the federated learning framework eliminates the need for centralized route computation, thereby reducing communication overhead and improving system responsiveness.

The detailed performance comparison results from 30 independent simulation runs with 1000-node topology, as presented in Table 5, provide quantitative evidence of the proposed protocol’s superiority across multiple performance dimensions. The table demonstrates that the proposed FL-Adaptive protocol achieves the highest throughput, lowest latency, and minimal packet loss rates while maintaining excellent scalability characteristics, with all metrics showing statistical significance at p < 0.01. The comparison reveals particularly significant improvements in packet loss rate, where the proposed protocol achieves less than 1% packet loss compared to 5–12% for traditional protocols, indicating superior reliability and robustness.

Table 5.

Protocol performance comparison results*.

Protocol Name	Throughput (Mbps)	Average Delay (ms)	Packet Loss Rate (%)	Scalability Score	Energy Efficiency
AODV	8.2 ± 0.7	125.3 ± 8.5	8.7 ± 0.6	6.2 ± 0.4	5.8 ± 0.5
DSR	7.8 ± 0.8	132.1 ± 9.2	12.4 ± 0.8	5.9 ± 0.5	5.5 ± 0.6
OLSR	9.1 ± 0.8	98.7 ± 7.1	6.3 ± 0.5	7.1 ± 0.5	6.2 ± 0.5
DSDV	6.9 ± 0.9	145.8 ± 10.3	9.8 ± 0.7	5.4 ± 0.6	5.2 ± 0.6
Centralized ML	12.5 ± 1.0	78.4 ± 5.6	3.2 ± 0.3	6.8 ± 0.4	7.3 ± 0.6
FL-Adaptive	16.8 ± 0.9	42.1 ± 3.2	0.8 ± 0.1	9.4 ± 0.3	9.1 ± 0.4

*Results from 30 independent simulation runs with 1000-node network topology. Values shown as mean ± standard deviation. All differences between FL-Adaptive and baseline protocols are statistically significant at p < 0.01.

The comprehensive performance analysis across different network scales and traffic conditions with error bars representing 95% confidence intervals from 30 independent runs, as illustrated in Fig. 4, demonstrates the robustness and scalability advantages of the proposed adaptive routing protocol. The figure reveals that while traditional protocols show significant performance degradation as network size increases, the proposed protocol maintains consistent performance improvements across all tested scenarios. The adaptive mechanism effectively handles network dynamics and device heterogeneity, resulting in stable performance even under challenging operational conditions.

Fig. 4.

Routing protocol performance metrics comparison analysis. All subplots include error bars (vertical lines through data points) representing 95% confidence intervals computed from 30 independent simulation runs. Legend consistently shows: Red solid with circles = FL-Adaptive, Blue dashed with squares = AODV, Green dotted with triangles = DSR, Orange dash-dot with diamonds = OLSR, Purple short-dash with stars = Centralized ML.

The scalability evaluation demonstrates that the proposed protocol maintains superior performance characteristics even as network size increases to 10,000 devices, while traditional protocols exhibit exponential performance degradation beyond 1000 devices⁴⁹. The distributed nature of the federated learning framework enables effective load distribution across edge nodes, preventing centralized bottlenecks that typically limit the scalability of conventional routing approaches. The adaptive mechanism continuously optimizes routing parameters based on current network conditions, ensuring consistent performance across different deployment scales.

A noteworthy observation from Fig. 3; Table 5 is that the federated learning approach outperforms the centralized ML baseline across multiple metrics, including convergence time, delay, and energy consumption. This result may appear counterintuitive since centralized training typically achieves equal or superior model quality compared to federated learning when given equivalent computational resources and complete data access. However, several factors specific to large-scale PIoT deployments explain this phenomenon. First, the centralized ML approach requires transmitting raw training data from all edge nodes to a central server, introducing substantial communication latency and bandwidth consumption that directly increases end-to-end delay and energy expenditure. In our 1000-node simulation, each edge node must upload approximately 2 MB of training data per round, creating a communication bottleneck at the central aggregator that does not exist in the federated setting where only 2.4 MB model parameters are exchanged regardless of local dataset size. Second, the central server becomes a single point of congestion: as routing queries queue at the overloaded central node, response latency increases dramatically, whereas federated learning allows each edge node to make local routing decisions immediately using its locally trained model. Third, federated learning enables faster adaptation to localized network changes because local models update every 30 s based on current observations, while centralized training introduces additional delay as data must first travel to the server, be processed, and updated models must then propagate back to edge nodes. Finally, in distributed PIoT environments where network conditions vary substantially across different geographic regions, local models can specialize to their respective neighborhoods, collectively achieving better real-time routing performance than a single global model that must generalize across all conditions. We acknowledge that in smaller networks with abundant bandwidth and low-latency central connectivity, centralized ML could potentially match or exceed federated learning performance; however, our experimental conditions specifically target the challenging large-scale scenarios where federated approaches demonstrate clear practical advantages.

To evaluate the federated learning component specifically, we assess the training convergence and model accuracy. The federated learning model training process spans 100 global communication rounds with each round involving local training at 50 edge nodes for 5 epochs each, model parameter upload to fog layer aggregators, global aggregation using FedProx algorithm, and updated model broadcast to all participants. Training loss decreases from 0.85 to 0.12 over 100 rounds while validation loss decreases from 0.92 to 0.18 and stabilizes after round 75, with convergence time approximately 50 min for 100 rounds at 30-second aggregation intervals and no significant overfitting observed. Model accuracy metrics demonstrate strong prediction capability: Mean Absolute Error (MAE) of 0.08 for traffic volume prediction, Root Mean Squared Error (RMSE) of 0.12 for link quality prediction, R² score of 0.89 for overall network state prediction, and 94.3% accuracy for binary congestion detection. Due to non-IID data distribution, individual edge nodes achieve varying performance with best client at 96% accuracy, worst at 88% accuracy, standard deviation of 3.2% across clients, and global model achieving 92.7% accuracy averaged across all test sets. The federated learning introduces manageable communication overhead with model size of 2.4 MB, requiring 2.4 MB upload and 2.4 MB download per client per round, totaling 240 MB per round for 50 clients, utilizing less than 2% of available network capacity, with communication needs reduced by 60% after round 50 as training converges and updates become less frequent.

The robustness analysis under various failure scenarios, including node failures, link disruptions, and traffic variations, confirms the superior fault tolerance of the proposed protocol. The federated learning component enables rapid adaptation to network changes through collaborative learning and distributed optimization, maintaining network connectivity and performance even during adverse conditions. The experimental results validate that the adaptive routing protocol provides significant advantages in terms of reliability, efficiency, and scalability, making it particularly suitable for large-scale PIoT deployments where traditional protocols fail to meet performance requirements.

Energy efficiency optimization effect analysis

The energy efficiency optimization algorithm demonstrates remarkable performance improvements under varying load conditions, achieving significant energy savings ranging from 45% to 65% compared to traditional power management approaches across different network traffic scenarios⁵⁰. Prior empirical studies provide important context for understanding these results. Research on embedded neural network deployment has shown that ARM Cortex-M4 and Cortex-A class processors can execute neural network inference with power consumption in the 1–5 W range, with energy consumption per inference operation measured in the millijoule scale^42–44. Studies on NVIDIA Jetson edge devices demonstrate that careful power management and model optimization can reduce DNN energy consumption by 11–30% compared to baseline execution strategies⁴³. Our experimental observations align with these findings: edge node power consumption during training operations averages 2.1 W with standard deviation of 0.3 W based on our simulation energy models calibrated against reported embedded device specifications.

While machine learning-based approaches incur computational overhead, the results show counterintuitively lower response times and energy consumption than traditional algorithms for several reasons. After initial training converges around round 75, model inference (forward pass) requires only 5-10ms compared to traditional algorithms that must compute routes from scratch for each request taking 50-150ms. The attention mechanism enables parallel processing of multiple routing requests, significantly reducing average response time under high load conditions. Although model training incurs upfront energy costs measured at approximately 120 J per edge node per training round, the optimized routing decisions reduce unnecessary retransmissions by 23% on average, avoid congested paths thereby decreasing packet delay energy, and enable better sleep scheduling for idle devices. Over the 5000-second simulation period, cumulative energy savings from optimized routing exceed training energy costs by a factor of 3–4×, validating the practical viability of our approach even when accounting for full machine learning overhead. The training energy overhead, amortized over the operational lifetime, represents only 15% of total system energy consumption, a cost that is more than offset by intelligent routing optimization gains. The experimental evaluation reveals that the proposed federated learning-driven optimization strategy effectively adapts to dynamic load variations by intelligently adjusting device duty cycles, optimizing transmission power levels, and implementing coordinated sleep-wake schedules that minimize unnecessary energy consumption while maintaining service quality requirements.

The power consumption distribution analysis indicates that the proposed optimization algorithm achieves substantial reductions in both communication and computational energy expenditure across heterogeneous PIoT devices. Under low-load conditions, the algorithm demonstrates exceptional energy savings of up to 65% through intelligent sleep scheduling and adaptive duty cycling mechanisms that coordinate device operations to minimize idle power consumption. During medium-load scenarios, the optimization strategy maintains energy savings of approximately 50% by implementing dynamic load balancing and predictive resource allocation that prevents energy wastage due to suboptimal task distribution.

The battery life extension evaluation reveals significant improvements in device operational longevity, with battery-powered sensors experiencing 2.5 to 3.8 times longer operational periods compared to baseline energy management approaches. The adaptive power management mechanism enables fine-grained control over device energy consumption patterns, allowing critical devices to prioritize essential operations while less critical devices enter extended sleep modes during low-activity periods. The federated learning component continuously optimizes energy allocation strategies based on historical usage patterns and device performance data, resulting in progressively improved battery life extension as the system learns optimal energy management policies. As illustrated in Fig. 5, the protocol demonstrates robust performance under various failure scenarios: under node failures ranging from 5% to 30% failure rates, FL-Adaptive maintains packet delivery ratio above 95% compared to 70–85% for traditional protocols; under link disruptions with varying interference levels from − 80dBm to -60dBm, FL-Adaptive sustains throughput degradation below 15% versus 40–60% for baseline methods; and after network partition events, FL-Adaptive achieves recovery time of 8.3 ± 1.2 s compared to 25–45 s for conventional protocols. Figure 6 presents detailed battery lifetime analysis: average device lifetime increases from 12.4 ± 1.8 days with static scheduling to 42.7 ± 3.9 days with FL-based optimization; lifetime distribution across heterogeneous device types shows temperature sensors achieving 65.3 ± 5.2 days, smart meters 58.1 ± 4.7 days, and voltage monitors 38.9 ± 3.6 days under the proposed optimization; and energy depletion curves demonstrate that FL-optimized devices maintain above 20% residual energy after 40 days while baseline approaches deplete to 0% within 15 days.

Fig. 5.

Robustness analysis under various failure scenarios. (a) Packet delivery ratio under node failure rates from 5 to 30%. (b) Throughput degradation under link disruptions with interference levels from − 80dBm to -60dBm. (c) Network recovery time after partition events. Red solid circles = FL-Adaptive, Blue dashed squares = AODV, Green dotted triangles = DSR, Orange dash-dot diamonds = OLSR, Purple short-dash stars = Centralized ML.

Fig. 6.

Battery lifetime extension analysis. (a) Average device lifetime comparison across optimization strategies. (b) Lifetime distribution across heterogeneous device types under FL-based optimization. (c) Energy depletion curves over 40-day operation period. Red solid circles = FL-based, Blue dashed squares = Static Scheduling, Green dotted triangles = Dynamic Load Balancing, Orange dash-dot diamonds = Adaptive Duty Cycling, Purple short-dash stars = Predictive Migration.

The comprehensive energy efficiency and power consumption analysis with error bars representing 95% confidence intervals from 30 independent runs, as presented in Fig. 7, demonstrates the superior performance of the proposed optimization algorithm across multiple evaluation metrics including total energy consumption, device-level power distribution, and battery lifetime extension. The figure reveals that the proposed approach achieves consistent energy efficiency improvements across all tested scenarios, with particularly significant gains observed in heterogeneous network environments where traditional approaches struggle to accommodate diverse device energy requirements. The analysis clearly illustrates the effectiveness of the federated learning-driven optimization in reducing overall system energy consumption while maintaining optimal network performance.

Fig. 7.

Energy efficiency optimization effects and power consumption analysis comparison. All subplots include error bars representing 95% confidence intervals from 30 independent simulation runs. Red solid circles = FL-based, Blue dashed squares = Static, Green dotted triangles = Dynamic Load, Orange dash-dot diamonds = Adaptive Duty, Purple short-dash stars = Predictive.

The feasibility analysis in realistic power grid environments confirms the practical applicability of the proposed energy optimization strategy, demonstrating stable performance improvements across diverse operational scenarios including peak load periods, equipment maintenance cycles, and emergency response situations⁵¹. The algorithm successfully adapts to varying network conditions and device capabilities, maintaining energy efficiency gains even under challenging operational constraints such as limited communication bandwidth, intermittent connectivity, and heterogeneous device performance characteristics.

The practical implementation evaluation reveals that the proposed optimization algorithm integrates seamlessly with existing power grid infrastructure, requiring minimal modifications to deployed PIoT devices while delivering substantial energy efficiency improvements. The distributed nature of the federated learning framework enables gradual deployment across network segments, allowing utilities to implement the optimization strategy incrementally without disrupting existing operations. The experimental results validate that the energy efficiency optimization algorithm provides significant practical benefits for real-world PIoT deployments, offering substantial energy savings, extended device lifetimes, and reduced operational costs while maintaining high service quality and system reliability.

Conclusion

This paper presents a novel adaptive routing protocol for large-scale Power Internet of Things based on edge computing and federated learning, addressing the critical challenges of heterogeneous device coordination, energy efficiency optimization, and scalable communication management in modern power grid infrastructure. The primary innovation lies in the integration of distributed edge computing capabilities with federated learning mechanisms to enable intelligent, privacy-preserving routing optimization that adapts dynamically to changing network conditions and device characteristics⁵².

The technical contributions include the development of a hierarchical system architecture that effectively coordinates heterogeneous PIoT devices through edge-based processing nodes, the design of a federated learning-driven dynamic routing optimization algorithm that enables collaborative learning without centralized data sharing, and the implementation of comprehensive energy efficiency optimization strategies that significantly reduce system power consumption while maintaining service quality requirements. The proposed protocol framework successfully addresses the interoperability challenges inherent in heterogeneous device environments through adaptive parameter adjustment mechanisms and intelligent path selection algorithms.

The experimental validation demonstrates substantial performance improvements compared to traditional routing protocols, achieving 35–50% higher network throughput, 40–60% reduction in end-to-end delay, and 45–65% energy savings across various network scales and operational conditions. The results confirm the protocol’s superior scalability characteristics, maintaining consistent performance even as network size increases to 10,000 devices, while traditional approaches exhibit significant performance degradation beyond 1000 devices.

The research limitations include the assumption of relatively stable network topology during federated learning convergence periods and the computational overhead associated with continuous model training at edge nodes⁵³. Future research directions encompass the integration of blockchain technology for enhanced security and trust management, the development of hybrid learning approaches that combine federated learning with reinforcement learning for improved adaptability, and the extension of the framework to support multi-objective optimization scenarios with conflicting requirements.

The technology application prospects are promising, with potential deployment in smart grid modernization projects, renewable energy integration systems, and industrial IoT applications where heterogeneous device coordination and energy efficiency are critical operational requirements⁵⁴. The proposed framework provides a foundation for next-generation power grid communication infrastructure that can adapt to evolving technological requirements while maintaining high performance, reliability, and energy efficiency standards.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

**Yong Zhang: ** Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing—original draft, Writing—review & editing, Visualization, Supervision, Project administration, Funding acquisition.**Sixiang Zhang: ** Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing—original draft, Visualization.**Shipeng Li: ** Software, Validation, Formal analysis, Investigation, Data curation, Writing—review & editing.**Bo Yang: ** Validation, Formal analysis, Investigation, Resources, Data curation, Writing—review & editing.**Peng Chen: ** Validation, Investigation, Resources, Data curation, Writing—review & editing. All authors have read and agreed to the published version of the manuscript.

Funding

No funding was received for conducting this study.

Data availability

The complete NS-3 extension implementation including custom modules for federated learning support is provided in the Supplementary Materials. This codebase comprises approximately 5,000 lines of C++ and Python code integrating PyTorch 1.12 via Python bindings, custom application-layer federated aggregation protocols, extended energy models for computational cost tracking, and synchronization protocols for model parameter exchange. The Supplementary Materials contain complete simulation scripts, model training code, configuration files, and documentation sufficient to reproduce all experimental results presented in this study.The simulation data and experimental results supporting the conclusions of this article are available from the corresponding author upon reasonable request. Due to the proprietary nature of some power grid infrastructure models, certain datasets may be subject to confidentiality agreements and cannot be publicly shared.

Declarations

Competing interests

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Ethics approval

This study does not involve human participants, animals, or any biological materials. The research is based on computer simulations and mathematical modeling of network protocols for Power Internet of Things systems. Therefore, ethics approval is not applicable for this work.