Dynamic routing with deterministic scheduling and self-healing recovery for industrial thermal storage networks in a hybrid physical emulation testbed

1 Introduction

1.1 Industrial thermal storage and communication requirements

Industrial thermal storage systems contribute to energy efficiency, grid flexibility, renewable-energy integration, and industrial decarbonization by shifting thermal demand and improving the coordination of heat generation, storage, and consumption (Nagaraju et al., 2022; Dass et al., 2023; Ogundile and Alfa, 2017). Their effective operation, however, depends not only on the physical performance of storage tanks, heat exchangers, pumps, and valves, but also on the reliability and timeliness of the supporting communication network.

Thermal-storage control requires the continuous exchange of temperature, pressure, flow, valve-position, pump-status, and thermal-state-of-charge information. These messages do not necessarily have equal operational importance. For example, a routine monitoring packet from a thermally stable storage unit may tolerate a moderate communication delay, whereas a temperature measurement approaching a safety threshold or a rapidly changing pressure signal may require immediate and highly reliable delivery. Communication priority should therefore reflect both network conditions and the current physical state of the thermal process.

Conventional and network-adaptive routing methods generally respond to communication-layer indicators such as hop count, link quality, congestion, or packet loss. They may therefore fail to prioritize thermally critical control traffic when network degradation and process disturbances occur simultaneously. Delayed or lost measurements can result in late control commands, unnecessary route changes, reduced thermal stability, and extended fault-recovery periods (Shen et al., 2017; Ogundile and Alfa, 2017). This motivates a routing mechanism in which communication decisions are directly informed by the operational state and urgency of the associated thermal assets.

The present study does not represent a complete industrial field deployment. It evaluates the proposed routing method in a hybrid physical–emulation Hardware-in-the-Loop (HIL) environment in which selected thermal and control components are physical, additional communication nodes and faults are emulated, and selected thermal dynamics are represented through simulation models.

1.2 Limitations of network-only adaptive routing

Existing routing approaches for wireless sensor networks and Industrial Internet of Things applications commonly select or update paths using network-related indicators such as hop count, expected transmission count (ETX), link quality, residual energy, queue occupancy, traffic congestion, and trust or security scores (Nagaraju et al., 2022; Hu et al., 2021). These indicators are important for communication efficiency and network lifetime, but they do not necessarily represent the operational urgency of the physical process associated with each data flow. In an industrial thermal-storage network, route-selection decisions may also need to account for temperature deviation from the desired set point, rate of temperature change, pressure deviation, flow instability, thermal state of charge, remaining safety margin, and overall process urgency. Many routing methods optimize communication performance without directly incorporating the instantaneous physical condition of the thermal process. Consequently, a network may select a route that is efficient according to conventional network metrics but unsuitable for a thermally critical control flow (Yun and Yoo, 2021; Han et al., 2022).

The resulting limitation is referred to in this study as the network–thermal state gap. It becomes particularly important during peak demand, congestion, intermittent connectivity, sensor degradation, and node or link failures. Addressing this gap requires a routing mechanism that combines thermal-process state information with communication-layer measurements and activates an explicitly defined self-healing procedure when service requirements are violated (Sah et al., 2023; Nguyen et al., 2018). Table 1 summarizes the positioning of representative routing studies and the proposed hybrid HIL framework. The table emphasizes whether the reviewed approaches incorporate physical thermal-state variables, deterministic communication support, and hybrid physical–emulation validation.

Table 1. Representative recent work and methodological positioning of the proposed thermal-state-aware framework.

1.3 Problem statement

The research problem is to determine how physical thermal criticality and dynamic network conditions can be combined in real-time route selection while satisfying latency, reliability, bandwidth, synchronization, and fault-recovery requirements in a hybrid physical–emulation industrial thermal-storage environment.

The routing controller must combine measurements originating from physical thermal assets, simulated thermal models, physical network devices, and emulated communication links without treating these evidence sources as equivalent. It must also avoid routes that violate flow-specific deadlines, reliability limits, or available-capacity constraints while responding to both physical and emulated faults.

Accordingly, the central optimization question addressed in this study is:

Which feasible communication route minimizes the combined network and process-related cost while maintaining the deadline, reliability, bandwidth, and recovery requirements of thermally critical flows?

This formulation enables the proposed routing method to be evaluated against technically defined static and network-state adaptive baselines under identical physical, emulated, and simulated operating conditions.

1.4 Research objectives

The objectives of this study are as follows:

• To formally define a thermal-state-aware routing and self-healing mechanism that combines communication-state and thermal-process variables.
• To develop a hybrid physical–emulation HIL framework connecting physical thermal equipment and controllers with network emulation and thermal simulation.
• To compare the proposed method with precisely defined static and network-state adaptive routing baselines under identical topology, traffic, load, and fault conditions.
• To evaluate physical measurements, integrated hybrid measurements, emulation-derived results, and simulation-derived results as separate evidence categories.
• To conduct statistically valid performance analysis using independently initialized experimental runs while accounting for temporal dependence in packet and sensor observations.
• To develop a scenario-based techno-economic and environmental assessment using documented capital-cost, operating-cost, energy-price, project-lifetime, discount-rate, and carbon-emission assumptions.

1.5 Contributions and scope

The principal contributions of this study are summarized as follows:

The study evaluates prototype-level performance within a controlled hybrid physical–emulation HIL environment and does not claim complete industrial field deployment, commercial validation, or realized long-term economic and environmental benefits.

• A formal routing model that combines thermal-process variables, including temperature deviation, pressure, flow, and thermal criticality, with communication-state variables such as delay, jitter, packet delivery, congestion, and link availability.
• A self-healing route-recovery mechanism with explicitly defined fault-detection thresholds, backup-path activation rules, switching conditions, and recovery-confirmation criteria.
• A hybrid physical–emulation HIL architecture that transparently separates physical measurements (PHY), network-emulation outputs (EMU), simulation-derived thermal outputs (SIM), and integrated hybrid measurements (HYB).
• A reproducible comparison with static and network-state adaptive routing baselines implemented under consistent topology, traffic, synchronization, load, and fault conditions.
• Independent-run statistical validation that distinguishes experimental replicates from temporally correlated packet-level and sensor-level observations.
• A scenario-based techno-economic and environmental projection based on documented experimental inputs, simulation outputs, deployment assumptions, and sensitivity analysis.

2 Literature review

2.1 Industrial communication, TSN, and adaptive routing

Industrial communication systems combine wired Industrial Ethernet, wireless sensor networks, edge controllers, and supervisory platforms to support monitoring and closed-loop control in cyber-physical environments (Rajulu et al., 2024; Yun and Yoo, 2021). Industrial Ethernet technologies provide high bandwidth, stable connectivity, and interoperability with programmable logic controllers and supervisory systems, whereas wireless communication provides deployment flexibility for distributed sensing and difficult-to-access assets. Wireless links, however, remain sensitive to electromagnetic interference, multipath propagation, physical obstructions, congestion, and time-varying channel conditions, which may increase delay, jitter, and packet loss (Rajulu et al., 2024; Dass et al., 2023).

Time-Sensitive Networking (TSN) introduces deterministic communication functions for industrial traffic. IEEE 802.1AS supports clock synchronization, IEEE 802.1Qbv enables time-aware queue scheduling, IEEE 802.1Qci provides per-stream filtering and policing, and IEEE 802.1Qbu supports frame preemption. These mechanisms should be distinguished from network-layer routing. Routing determines the path followed by a data flow, whereas TSN functions control queue assignment, transmission timing, synchronization, traffic shaping, and stream protection along the selected path. Recent experimental and optimization studies have shown that IEEE 802.1Qbv performance depends strongly on gate-control-list synthesis, traffic assumptions, topology, and implementation-specific timing behavior (Xue et al., 2024; Torres-Macías et al., 2024; Ulbricht et al., 2024). Accordingly, thermal-state-aware path selection and deterministic TSN scheduling perform complementary functions in the proposed framework.

Software-Defined Networking (SDN) further supports industrial communication by separating forwarding from control and enabling centralized collection of link-quality, congestion, and topology information (Ahmed et al., 2018). Recent SDN-enabled IIoT work has examined controller selection, centralized traffic engineering, reliable route computation, and performance optimization across topologies with different node and edge counts (Ali et al., 2024). These studies support the use of common controller interfaces and explicit routing objectives, but they generally optimize communication-state variables without directly representing thermal-process urgency.

Energy-aware routing methods use residual energy, transmission cost, and node lifetime to avoid premature network partitioning (Nagaraju et al., 2022; Hu et al., 2021; Nguyen et al., 2018). Delay-aware methods prioritize paths with lower end-to-end latency or expected transmission count, while congestion-aware approaches incorporate queue occupancy, bandwidth availability, and packet-delivery probability (Yun and Yoo, 2021; Sah et al., 2023). Fault-tolerant routing and industrial control extend these methods through failure detection, alternative-path generation, controller reconfiguration, and recovery verification when nodes, links, sensors, or actuators become unavailable (Sharma et al., 2023; Han et al., 2022; Kalel and Singh, 2024).

Real-time network emulation provides an intermediate validation environment between numerical simulation and full industrial deployment. It allows physical devices to exchange real packets through configurable virtual nodes, links, queues, traffic loads, and fault conditions. Recent NS-3 real-time-emulation work emphasizes wall-clock synchronization, emulation lag, and implementation overhead as explicit validity constraints (Ottens et al., 2025). Such emulation is particularly useful for evaluating routing scalability and resilience while retaining interaction with physical controllers and sensing equipment.

2.2 Thermal energy management and communication requirements

Industrial thermal-storage systems exhibit slow but strongly coupled physical dynamics involving thermal inertia, stratification, charge and discharge cycles, pressure regulation, flow control, and heat-transfer losses (Kumbhar and Naik, 2022; Xu et al., 2021). Thermal inertia can temporarily reduce sensitivity to communication delay; however, persistent or poorly timed control delays may increase temperature-tracking error, disturb stratification, and reduce storage efficiency. The communication system must therefore support both periodic monitoring traffic and event-driven control messages.

Thermal state of charge provides an estimate of the available stored thermal energy and influences charging, discharging, and load-allocation decisions. Temperature, pressure, and flow measurements are used to determine whether pumps, valves, and heat exchangers are operating within their required limits. These measurements are commonly generated at different sampling rates because rapidly changing control variables require more frequent updates than slowly varying supervisory indicators.

Communication priority should consequently vary according to the physical condition of the associated thermal asset. A routine monitoring flow from a stable tank can tolerate a longer delay, whereas a flow carrying data from a rapidly increasing tank temperature, a pressure deviation, or a reduced safety margin should receive higher routing and scheduling priority. Delayed delivery of thermally critical information may result in late valve actuation, incorrect pump-speed adjustment, increased temperature overshoot, or prolonged recovery following a process disturbance.

This relationship creates a cross-domain control requirement: communication decisions should consider both network-state indicators and thermal-process urgency. A routing method based only on hop count, link quality, or residual energy may select an acceptable network path without recognizing that the corresponding packet is associated with a safety-critical or time-sensitive thermal condition.

2.3 Hybrid HIL, network emulation, and digital-twin validation

Validation approaches for industrial cyber-physical systems differ according to the degree of physical and virtual integration. A physical experiment evaluates components through actual hardware and measured operating conditions. Software-in-the-Loop testing evaluates control software against a fully simulated plant, while Hardware-in-the-Loop testing places physical sensors, controllers, or actuators in a closed loop with real-time simulation models. Co-simulation synchronizes two or more specialized simulation environments so that each model represents a different system domain. Correct synchronization, bounded exchange delay, common time bases, and explicit real-time constraints are therefore central to credible HIL and co-simulation experiments.

Network emulation differs from conventional simulation because real physical endpoints exchange packets through virtualized links and nodes whose delay, loss, bandwidth, congestion, and failure characteristics are controlled in real time. NS-3-based real-time emulation has been used to assess communication protocols under repeatable traffic and fault conditions while preserving interaction with physical devices (Ottens et al., 2025). OPC UA supports structured exchange of industrial process variables, whereas MQTT provides lightweight publish–subscribe communication for telemetry and supervisory data (Liu and Bellot, 2019). Recent digital-twin platforms also emphasize continued use of validated virtual commissioning models during operation, service testing, and process optimization (Chen et al., 2024).

A digital twin is a continuously updated virtual representation of a physical asset or process. In thermal-storage applications, digital twins may reproduce tank temperature, thermal stratification, flow behavior, heat loss, and state-of-charge dynamics. Their reliability depends on calibration against physical measurements and validation using indicators such as temperature root-mean-square error, flow error, pressure error, and temporal correlation (Lv et al., 2023; Han et al., 2022). Recent experimentally calibrated thermal-storage digital twins further demonstrate the importance of measured charging, discharging, simultaneous-operation, and natural heat-loss cycles when validating storage models (Rosato et al., 2025).

Fault-tolerant industrial-control studies commonly use reconfiguration, redundant sensing, alternative control laws, or smooth controller switching to preserve service after sensor or actuator faults (Kalel and Singh, 2024). These approaches are complementary to communication-route recovery: process-level controller reconfiguration addresses control-component faults, whereas the present routing method addresses path feasibility, traffic priority, and communication restoration.

The present study uses a hybrid physical–emulation HIL configuration. Selected thermal assets, sensors, programmable logic controllers, actuators, and communication devices form the physical subsystem. Additional communication nodes, links, traffic conditions, congestion, and failures are represented through real-time network emulation. MATLAB/Simulink and CFD-based models represent selected thermal processes and digital-twin states. Physical and virtual components exchange synchronized information through OPC UA and MQTT interfaces, allowing the routing method to be evaluated under repeatable normal, peak-load, and fault conditions.

2.4 Research gap analysis

Existing studies commonly address energy-aware routing, deterministic industrial communication, thermal-process optimization, HIL validation, or fault-tolerant control as separate research problems. Energy-aware and adaptive routing methods primarily optimize communication variables such as residual energy, link quality, congestion, expected transmission count, and trust (Nagaraju et al., 2022; Hu et al., 2021; Yun and Yoo, 2021; Han et al., 2022; Sah et al., 2023; Nguyen et al., 2018). TSN research develops scheduling and traffic-shaping mechanisms for deterministic service (Xue et al., 2024; Torres-Macías et al., 2024; Ulbricht et al., 2024), while SDN-based industrial networking provides centralized topology and traffic control (Ahmed et al., 2018; Ali et al., 2024). Thermal-management and digital-twin studies generally focus on process fidelity, storage control, or model validation (Lv et al., 2023; Rosato et al., 2025), and fault-tolerant control studies focus on maintaining process operation after component faults (Kalel and Singh, 2024).

Limited integrated evidence is available for a formally defined routing method that directly incorporates real-time thermal criticality into feasible-path selection while also using deterministic queue scheduling, explicit communication-fault recovery rules, and a transparently documented hybrid physical–emulation HIL evaluation. The contribution claimed in this study is therefore not a new artificial-intelligence model, a new TSN standard, or a complete industrial deployment. It is the integration and controlled evaluation of: (i) thermal-criticality calculation, (ii) constrained multi-criteria route selection, (iii) deterministic traffic-class assignment, and (iv) threshold-based self-healing recovery within the evaluated HIL configuration.

Table 2 summarizes this methodological positioning.

Table 2. Methodological gap matrix for routing, deterministic communication, digital twins, and hybrid HIL validation.

The proposed work addresses this limited integration by coupling a formal thermal-state-aware path-selection method with deterministic queue scheduling and an explicit self-healing mechanism. Its evaluation separates physical measurements, network-emulation outputs, simulation-derived thermal states, integrated hybrid measurements, derived metrics, and scenario-based economic projections within one reproducible experimental framework.

3 Proposed thermal-state-aware self-healing routing model

3.1 Overview

This study introduces a Thermal-State-Aware Dynamic Routing (TSDR) method that combines thermal-process criticality with real-time communication-state information for industrial thermal-storage networks. The method is deterministic and rule-based. It does not contain a trained artificial-intelligence, machine-learning, or reinforcement-learning component.

The TSDR controller receives temperature, pressure, flow, thermal state of charge, link delay, jitter, packet-delivery probability, queue utilization, bandwidth, and link-availability information. It generates a limited set of candidate paths, removes paths that violate capacity, delay, reliability, or availability constraints, and assigns a cost to every remaining path. The feasible path with the minimum cost is selected, after which the associated traffic is mapped to an appropriate deterministic communication queue.

The communication system is represented by the graphwhere the node set is partitioned asHere, VPHY contains physical sensors, programmable logic controllers, switches, gateways, and controllers, while VEMU contains the additional sensor, relay, and routing nodes represented through real-time network emulation. Similarly, the thermal-asset set is defined aswhere HPHY denotes physical thermal assets and HSIM denotes thermal assets or operating conditions represented through simulation models.

G=(V,E),(1)

V=VPHY∪VEMU,VPHY∩VEMU=∅.(2)

H=HPHY∪HSIM,HPHY∩HSIM=∅,(3)

Rerouting is initiated when a path becomes unavailable, violates a flow-specific service constraint, or remains degraded for a predefined number of monitoring intervals. Physical endpoints exchange data with the emulated network through a real-time gateway, while physical and simulated thermal states are synchronized through OPC UA and MQTT interfaces. Further details are included in Algorithm 1.

3.2 System, flow, and state model

Let F denote the set of periodic monitoring and event-driven control flows. For each flow f∈F, the model defines a source sf, destination df, origin type, packet size lf, packet rate rf, required bandwidth bf, maximum delay Dfmax, minimum delivery probability Qfmin, thermal criticality Kf, and traffic class cf. The set of candidate paths available to flow f is denoted by Pf.

The physical and emulated link set is partitioned asEach link e∈E is characterized by capacity Ce, delay de, jitter je, delivery probability qe, utilization ue, available bandwidth be, energy cost εe, availability ae, and fault state ϕe.

E=EPHY∪EEMU.(4)

Table 3 summarizes the principal notation used in the proposed model.

Table 3. Principal notation for the proposed routing model.

3.3 Thermal-state-aware route optimization

3.3.1 Thermal-state vector

For each thermal asset i, the thermal-state vector is defined aswhereis the temperature deviation from the required set point, andis the temperature-change rate.

sith(t)=[Ti(t),ΔTi(t),T˙i(t),Pi(t),Fi(t),SOCi(t)],(5)

ΔTi(t)=Ti(t)−Tiset(6)

T˙i(t)=Ti(t)−Ti(t−Δt)Δt(7)

Temperature, pressure, and flow are obtained directly from physical sensors for HPHY. Corresponding variables for HSIM are generated by the synchronized thermal model. Thermal state of charge is estimated from the available stored thermal energy and the defined operating limits.

3.3.2 Network-state vector

For every physical or emulated link e, the communication state is represented asPhysical link information is collected from switch, gateway, and controller logs. Emulated link information is obtained from NS-3 traces, while integrated end-to-end delay and delivery measurements are calculated using synchronized timestamps at the physical sender and receiver.

senet(t)=[de(t),je(t),qe(t),ue(t),be(t),ae(t)].(8)

3.3.3 Thermal criticality

The thermal criticality associated with asset i is calculated aswhere M^i(t)∈[0,1] is the normalized remaining safety margin. The weights satisfyThe calibration configuration usesThe score Ki(t) is bounded between zero and one. Flows with Ki(t)≥0.70 are classified as critical, flows with 0.40≤Ki(t)<0.70 are classified as priority, and the remaining flows are classified as routine monitoring traffic.

Ki(t)=clip[0,1]⁡[wT|ΔTi(t)|ΔTmax+wR|T˙i(t)|T˙max+wP|Pi(t)−Piset|ΔPmax+wF|Fi(t)−Fiset|ΔFmax+wM(1−M^i(t))],(9)

wT+wR+wP+wF+wM=1.(10)

wT=0.35,wR=0.20,wP=0.15,wF=0.15,wM=0.15.

The state is updated every 100 ms. A missing value is replaced by the most recent valid observation for a maximum of 300 ms. When the value remains unavailable beyond this interval, the associated flow is assigned Ki(t)=1 until valid sensing is restored.

3.3.4 Route-cost function

For flow f and candidate path p, the composite route cost iswhere:

Cf,p(t)=Kf(t)[αD^p(t)+βJ^p(t)+γ(1−Q^p(t))]+δU^p(t)+ηE^p(t)+ζR^p(t),(11)

The path-level quantities are calculated asandThe calibrated route-cost weights areThe selected path is

• D^p is the normalized end-to-end path delay;
• J^p is the normalized path jitter;
• Q^p is the path-delivery probability;
• U^p is the maximum utilization along the path;
• E^p is the normalized communication-energy cost;
• R^p is the normalized fault-risk score.

Dp(t)=∑e∈pde(t),Qp(t)=∏e∈pqe(t),(12)

Up(t)=maxe∈pue(t).(13)

α=0.30,β=0.15,γ=0.25,δ=0.15,η=0.05,ζ=0.10.

pf⋆(t)=arg⁡minp∈PfCf,p(t).(14)

3.3.5 Routing constraints

The path assignment is represented by the binary variableEach flow is assigned to one feasible path:The selected path must satisfy the delay and reliability requirements:The aggregate flow rate on each link must remain within its available capacity:A path containing an unavailable or confirmed failed link cannot be selected:

xf,p={1,if flowfuses pathp,0,otherwise.

∑p∈Pfxf,p=1,∀f∈F.(15)

Dp(t)≤Dfmax,(16)

Qp(t)≥Qfmin.(17)

∑f∈F∑p∈Pf∈pbfxf,p≤Ce,∀e∈E.(18)

xf,p=0if∃e∈psuch thatae(t)=0.(19)

3.3.6 Source-aware state fusion

Every state variable is stored together with a source labelwhere PHY denotes physical measurements, EMU denotes network-emulation outputs, SIM denotes thermal-simulation outputs, and HYB denotes measurements obtained across integrated physical and emulated paths.

σ(z)∈{PHY,EMU,SIM,HYB},(20)

Measurements are aligned to a common timestamp before route calculation. Physical thermal measurements and simulated thermal states are processed using the same normalization ranges, while physical and emulated communication metrics retain separate source identifiers in the experimental logs.

3.4 Functional modules and interfaces

The proposed architecture consists of the following coordinated modules:

Figure 1 illustrates the interaction among the physical, emulated, and simulated parts of the routing framework, while Figure 2 represents the timing and data-flow sequence for the hybrid HIL experiment.

• Thermal acquisition module: Collects physical temperature, pressure, and flow measurements and receives corresponding thermal states from the synchronized simulation models.
• State-awareness module: Constructs the thermal and network-state vectors, applies source labels, handles stale observations, and recalculates thermal criticality every 100 ms.
• State-aware multi-criteria route-optimization engine: Generates three loop-free candidate paths using a k-shortest-path procedure, checks their feasibility, evaluates their route costs, and selects the minimum-cost path.
• Deterministic communication layer: Applies path selection separately from TSN functions. IEEE 802.1AS provides time synchronization, IEEE 802.1Qbv controls time-aware queue scheduling, and IEEE 802.1Qci applies per-stream filtering and traffic policing.
• Dynamic control and feedback module: Promotes a flow to a higher traffic class when the temperature deviation exceeds 2∘C, pressure deviation exceeds 5%, flow deviation exceeds 10%, or thermal criticality exceeds 0.70.
• Supervisory analytics interface: Operates as physical software on the edge workstation and provides SCADA visualization, event logging, route-status monitoring, and evidence-source identification.

3.5 Fault detection and self-healing route recovery

The routing controller transmits heartbeat messages at 500 ms intervals and evaluates link status over consecutive monitoring windows. A physical or emulated link is classified as degraded when one of the following conditions is satisfied:or three consecutive heartbeat messages are missed. A failure is confirmed after three consecutive violations, corresponding to a maximum confirmation interval of 1.5 s.

qe(t)<0.95,(21)

de(t)>Dethr,(22)

For every active route, two feasible backup paths are precomputed. After failure confirmation, affected flows are transferred to the lowest-cost available backup path. A route-hysteresis interval of 2 s prevents repeated switching between paths.

Fault-recovery time is defined aswhere tfault is the time of fault injection or detection and tstable is the first instant at which:for five consecutive monitoring windows.

Trec=tstable−tfault,(23)

Qp(t)≥Qfmin,Dp(t)≤Dfmax,|ΔTi(t)|≤2∘C,(24)

Physical port or device failures are identified through physical link-state and heartbeat information. Emulated failures are identified from NS-3 link events and packet statistics. Simulated thermal disturbances do not directly remove a communication link; instead, they increase thermal criticality and trigger traffic reprioritization or rerouting when the current path no longer satisfies the required service constraints.

The primary path is restored only when it remains feasible for 10 consecutive monitoring windows and its cost is at least 10% lower than the active backup path.

3.6 Routing and self-healing algorithm

Algorithm 1Thermal-State-Aware Self-Healing Routing. Require: Physical sensor data, simulated thermal states, physical and emulated link metrics Ensure: Feasible route and deterministic traffic-class assignment 1: Read physical temperature, pressure, flow, and controller states 2: Read simulated thermal states and digital-model outputs 3: Collect physical and emulated delay, jitter, PDR, utilization, and availability 4: Align all observations using synchronized timestamps 5: Normalize thermal and communication variables 6: for all flows f∈F do 7: Compute thermal criticality Kf(t) 8: Generate k=3 loop-free candidate paths Pf 9: Remove paths violating delay, reliability, capacity, or availability constraints 10: for all feasible paths p∈Pf do 11: Compute route cost Cf,p(t) 12: end for 13: Select pf⋆(t)=arg⁡minCf,p(t) 14: Assign the corresponding traffic class and TSN queue 15: end for 16: Monitor heartbeats, link quality, and process-criticality thresholds 17: if a physical or emulated fault is confirmed then 18: Activate the minimum-cost feasible backup path 19: Apply the route-hysteresis interval 20: Verify recovery for five consecutive monitoring windows 21: end if 22: Update the route, fault, and evidence-source logs

3.7 Weight selection and sensitivity analysis

The thermal-criticality and route-cost weights were selected using a calibration dataset that was generated before final testing and was excluded from all reported test comparisons. The calibration dataset contained 48 independently initialized runs spanning normal operation, peak thermal demand, communication congestion, physical-link failure, emulated packet-loss, and simulated thermal-disturbance conditions, with eight repetitions conducted for each condition. Calibration used emulation seeds 11–18; the final test campaign used seeds 41–52, and no seed was shared between calibration and final testing.

All thermal and communication components were normalized to the interval [0,1] using engineering limits fixed before grid search. Candidate weights were generated on a 0.05 grid subject to the constraints that each weight remained between 0.05 and 0.50 and that the weights within each objective summed to one. After infeasible combinations were removed, 3,246 thermal-criticality vectors and 10,872 route-cost vectors were evaluated. A two-stage calibration procedure was used to limit computational cost. The thermal-criticality vectors were first evaluated using a fixed balanced route-cost configuration, after which the route-cost vectors were evaluated using the best-performing thermal-criticality vector. Each candidate was applied to the same calibration scenarios, traffic profiles, fault schedules, and initial conditions.

The calibration objective minimizedwhere the hatted terms denote min–max normalized calibration outcomes. The objective coefficients were λ1=0.30, λ2=0.25, λ3=0.30, and λ4=0.15. These coefficients assigned equal emphasis to tail latency and recovery performance, followed by packet-delivery reliability and thermal-tracking quality. When two configurations differed by less than 0.005 in Jcal, the configuration with the lower route-switch frequency was selected; remaining ties were resolved in favor of the lower average route-update time.

Jcal=λ1D^95+λ2(1−Q^)+λ3T^rec+λ4RMSE^T,(25)

The selected thermal-criticality weights were wT=0.35, wR=0.20, wP=0.15, wF=0.15, and wM=0.15. The selected route-cost weights were α=0.30, β=0.15, γ=0.25, δ=0.15, η=0.05, and ζ=0.10. The selected configuration achieved Jcal=0.214, compared with 0.238 for the equal-weight configuration and 0.227 for the second-ranked candidate. Its mean route-switch frequency was 3.8 switches per hour, while its mean route-update time was 7.6 ms. The final test runs were not inspected during calibration or tie resolution.

The criticality thresholds 0.40 and 0.70 were established from predefined engineering operating bands and subsequently verified during calibration. A score below 0.40 consistently represented thermally stable operation within the permitted temperature, pressure, flow, and safety-margin limits. Scores from 0.40 to below 0.70 represented emerging process deviation requiring prioritized communication, whereas scores of 0.70 or above corresponded to conditions requiring critical-traffic treatment. Threshold variations of ±0.05 were additionally examined during calibration, and the selected values provided the best balance between critical-event detection and unnecessary priority promotion.

Sensitivity analysis varied each thermal-criticality and route-cost weight independently by ±20%, followed by renormalization of the affected weight vector. Each perturbed configuration was evaluated using the final fixed test protocol, and changes in P95 latency, PDR, recovery time, temperature RMSE, and route-switch frequency were recorded. Across all perturbations, P95 latency varied by 2.1%–6.4%, PDR by 0.1–0.6 percentage points, recovery time by 3.2%–8.7%, temperature RMSE by 1.8%–5.9%, and route-switch frequency by 4.5%–12.6%. The relative performance ranking of the proposed method and the three baselines remained unchanged across all evaluated perturbations. The corresponding results are reported in Section 6.7.

3.8 Computational complexity

Candidate paths are generated using a k-shortest-path procedure with k=3. For one flow, the dominant computational complexity isFor |F| simultaneous flows, the route-update complexity isThe routing controller operates at a 100 ms update interval. On an eight-core 3.2 GHz edge controller with 16 GB RAM, the average route-update time is 7.8 ms, the observed worst-case time is 21.4 ms, and the routing process uses approximately 220 MB of memory. These values remain below the configured update deadline for the evaluated hybrid topology.

O(k|E|log⁡|V|).(26)

O(|F|k|E|log⁡|V|).(27)

3.9 Architecture-to-metric mapping

Table 4 links each architectural component to its decision function, output, evaluation metric, and evidence source.

Table 4. Architecture-to-metric mapping for the proposed framework.

4 Hybrid physical–emulation hil methodology and experimental protocol

4.1 Experimental design and evidence domains

The experimental design evaluates the effect of routing architecture on communication reliability, fault recovery, and thermal-control performance within a controlled hybrid physical–emulation Hardware-in-the-Loop (HIL) environment. The independent variable is the routing configuration, comprising static shortest-path routing, network-state adaptive routing, TSN-enabled network-state adaptive routing, and the proposed Thermal-State-Aware Dynamic Routing (TSDR) method.

The experimental factors include normal operation, peak thermal demand, communication congestion, physical node and link failures, emulated packet loss and intermittent connectivity, and simulated sensor or thermal-process disturbances. The primary outcomes are the 95th-percentile end-to-end latency and fault-recovery time. Secondary outcomes include jitter, Packet Delivery Ratio (PDR), deadline-miss ratio, temperature-tracking root-mean-square error, Energy Utilization Factor (EUF), Coefficient of Performance (COP), controller overhead, and electrical power consumption.

One independently initialized HIL run constitutes the experimental unit. Each run begins with restored routing tables, reset controller states, defined thermal initial conditions, and a unique emulation seed. Packet-level and sensor-level observations are treated as temporally nested measurements rather than independent experimental replicates as mentioned in Table 5.

Table 5. Main physical HIL components.

The complete equipment inventory is provided in Supplementary Table S23.

The evidence generated by the framework is classified as follows: The design evaluates three testable principles. First, bidirectional state awareness is assessed by determining whether thermal criticality changes route selection and whether communication degradation affects thermal-control quality. Second, multi-timescale operation is evaluated through 100 ms routing updates, 10 ms PLC control cycles, and slower thermal-energy supervisory intervals. Third, graceful degradation is assessed by measuring whether the system maintains defined communication and thermal-service limits during partial physical or emulated failures.

• PHY: measurements obtained directly from physical sensors, PLCs, actuators, switches, controllers, and power meters;
• EMU: link, queue, traffic, topology, and fault information generated by the real-time network emulator;
• SIM: thermal states generated by MATLAB/Simulink and reduced-order CFD models;
• HYB: integrated measurements obtained when physical endpoints communicate through emulated network paths;
• DER: metrics calculated from PHY, EMU, SIM, or HYB observations;
• EST: scenario-based economic and environmental projections.

4.2 Physical subsystem

The physical subsystem contains three 500 L thermocline storage tanks filled with a 70:30 water–propylene-glycol mixture and operated between 20∘C and 85∘C. Each tank is instrumented with eight Class A PT100 resistance-temperature sensors positioned at equally spaced vertical levels to capture thermal stratification.

The complete physical instrumentation includes 45 measurement devices: 24 PT100 temperature sensors, six Endress+Hauser Promag W 400 electromagnetic flowmeters, six Cerabar PMC51B pressure transmitters, three level transmitters, and six Siemens SENTRON PAC3200 power meters. Temperature sensors were calibrated at 25∘C, 55∘C, and 80∘C, while flow and pressure devices were verified using zero-, mid-, and full-scale calibration points before the experimental campaign.

Two Siemens S7-1500 CPU 1516-3 PN/DP PLCs execute the thermal-control logic with a 10 ms scan cycle. Six Danfoss AB-QM modulating valves regulate the hydraulic branches, while three Grundfos TPE3 pumps, each rated at 2.2 kW, control circulation through three 25 kW plate heat exchangers. Physical communication is supported by two Cisco IE4000 industrial Ethernet switches and two Siemens SCALANCE W788-2 wireless access points.

The supervisory and routing software executes on an eight-core 3.2 GHz edge controller with 16 GB RAM and Ubuntu 22.04. Electrical power is measured at 1 s intervals using the physical power meters positioned at the pumps, controllers, and communication equipment as described in Tables 6, 7.

Table 6. Integrated physical and emulated network configuration.

Table 7. Traffic classes and communication requirements.

4.3 Network-emulation subsystem

The network-emulation subsystem is implemented in NS-3.41 using the real-time scheduler. It represents 146 logical nodes, comprising 132 virtual sensing endpoints, 12 relay or routing nodes, and two virtual aggregation nodes. Emulated wired links operate between 100 Mb/s and 1 Gb/s with propagation delays of 0.5–2 ms. Wireless links follow IEEE 802.11n with data rates between 6 and 54 Mb/s.

A log-distance propagation model combined with Nakagami fading represents industrial wireless attenuation. Packet losses are generated using a Gilbert–Elliott burst-loss model, while congestion is produced through Poisson background traffic and configurable on–off bursts. Four-band priority queuing is used to represent the deterministic traffic classes. Network-state information, including delay, jitter, delivery probability, queue occupancy, bandwidth, and availability, is exported to the routing controller every 100 ms.

The emulation runs use seeds 41–52, with one seed assigned to each independent repetition. Physical packets enter and leave the emulated topology through Linux TAP interfaces connected to the edge gateway. The static, adaptive, and proposed routing algorithms are implemented through the same controller interface to ensure equivalent packet-processing and logging paths.

4.4 Thermal simulation and digital-twin subsystem

The thermal simulation is implemented in MATLAB R2025b and Simulink using a fixed-step fourth-order solver. The internal integration step is 10 ms, while state exchange with the physical and emulated subsystems occurs every 100 ms. The model represents tank energy balance, heat loss, inlet mixing, charge/discharge dynamics, valve position, pump flow, and thermal state of charge.

A reduced-order CFD model represents multi-zone vertical stratification using twelve thermal layers per virtual tank. The simulated subsystem contains six additional thermal-storage units and configurable demand profiles derived from nominal, peak, and disturbance conditions. Physical inlet temperature, outlet temperature, pressure, flow, and valve-position measurements are used for parameter calibration.

Model calibration minimizes the difference between measured and simulated temperature and flow trajectories. Validation against an independent physical cycle produced a temperature RMSE of 0.42∘C, a flow mean absolute error of 0.28Lmin−1, and a pressure mean absolute error of 2.1 kPa. These validated models provide the SIM thermal variables used by the routing controller.

4.5 Physical–emulation and physical–simulation interfaces

Physical packets are transferred to the NS-3 real-time environment through a TAP bridge attached to the industrial edge gateway. Packets are timestamped immediately before transmission and after delivery using hardware-assisted timestamps. The real-time scheduler maintains wall-clock execution, and a run is automatically restarted when accumulated emulation lag exceeds 2 ms.

Physical process measurements are published to the simulation layer through OPC UA at 10 Hz. Simulated tank states, demand variables, and derived thermal criticality values are returned using the same exchange interval. MQTT with QoS 1 is used for supervisory telemetry and event messages, whereas time-critical control variables use OPC UA PubSub.

Values are encoded using SI units and aligned to a common timestamp. Linear interpolation is applied to one missing simulation interval, while the most recent valid physical observation is retained for a maximum of 300 ms. Control commands generated by the supervisory model are transmitted to the PLCs through OPC UA and are subjected to local physical safety interlocks before actuation.

4.6 Timing and synchronization

The Cisco IE4000 grandmaster provides IEEE 802.1AS generalized Precision Time Protocol synchronization to the physical switches, PLC gateways, edge controller, and network-emulation host. Hardware timestamping provides physical clock alignment within ±50μs, while the maximum accepted PHY–EMU–SIM alignment error is 0.5 ms.

Packet events are logged at transmission, queue entry, queue exit, routing decision, and reception. Thermal states are logged every 100 ms, PLC events every 10 ms, and electrical-power measurements every 1 s. Clock-offset estimates are updated every 1 s, and linear drift compensation is applied during post-processing.

4.7 Integrated network topology

The integrated hybrid topology contains 14 physical network endpoints and 146 emulated nodes, giving a total logical topology of 160 nodes. The physical endpoints comprise two PLCs, two industrial switches, two wireless access points, three sensor gateways, two protocol gateways, one edge controller, one SCADA workstation, and one data-logging server.

The topology includes 12 emulated relay nodes, two emulated aggregation nodes, two physical gateways, and three physical controllers. The average end-to-end path length is 3.6 hops, with a maximum of seven hops. Physical wired links operate at 1 Gb/s, physical wireless links at up to 72.2 Mb/s, and emulated links between 6 Mb/s and 1 Gb/s.

Scalability beyond the physical endpoint count is evaluated through emulation while preserving physical sensing, PLC control, timestamping, and actuation at the boundaries of the integrated topology.

4.8 Deterministic communication and TSN configuration

The physical deterministic communication layer implements IEEE 802.1AS synchronization, IEEE 802.1Qbv time-aware scheduling, IEEE 802.1Qci per-stream filtering and policing, and IEEE 802.1Qbu frame preemption. Four traffic classes are mapped to Priority Code Point values 7, 5, 3, and 0.

The gate-control cycle is 1 ms and contains a 250μs safety window, a 350μs closed-loop control window, a 250μs monitoring window, a 100μs background window, and a 50μs guard band. Critical streams are policed at twice their nominal packet rate, and lower-priority frames may be preempted when safety or control traffic becomes active. Queue and routing priorities are updated every 100 ms.

The 1 ms gate cycle represents local deterministic switching and is distinct from application-level end-to-end latency. End-to-end latency includes physical protocol handling, emulated propagation and queueing, gateway processing, and delivery to the destination controller.

4.9 Traffic model

Four traffic classes are used throughout the experiments. Safety and control flows originate from both physical and simulated thermal assets, while monitoring and background flows include physical, simulated, and emulated sources.

Background traffic follows a Poisson process during nominal operation and an on–off burst process during congestion experiments. OPC UA process variables are sampled every 100 ms, while MQTT monitoring messages are published every 1 s.

4.10 Routing baselines

4.10.1 Common implementation controls

All four routing configurations used the same physical equipment, integrated source–destination flow set, topology, traffic profiles, thermal initial conditions, fault schedules, random seeds, run durations, packet-processing interface, timestamping procedure, and logging infrastructure. Link metrics were calculated over a 1 s sliding measurement window that was updated every 100 ms and normalized using the same limits listed in Table 8 and the network_metric_limits.yaml configuration file. Adaptive methods updated route state every 100 ms. Candidate generation used the same loop-free k-shortest-path implementation with k=3, and route installation used the same controller-to-switch interface with a measured mean installation delay of 8.9 ms and a P95 installation delay of 14.7 ms. Each run included the same 10 min warm-up period and scenario-specific measurement interval. TSN was enabled only for the two explicitly identified TSN configurations.

Table 8. Measurement boundaries and evidence sources for the principal outcomes.

4.10.2 Static shortest-path routing

Static routing assigned unit cost to every available link,and selected the feasible path with the minimum hop count. Equal-hop ties were resolved first by lower configured propagation delay and then by lexicographic node identifier. Paths were not recomputed for congestion, jitter, packet loss, or thermal state. Route recomputation occurred only after confirmed connectivity loss using the common physical/emulated availability signal. Static routing used a single first-in–first-out DropTail queue with a maximum capacity of 256 packets per egress interface and no PCP-based traffic differentiation; no gate-control list, precomputed backup path, thermal promotion, or route hysteresis was applied.

Cpstatic=∑e∈p1,(28)

4.10.3 Network-state adaptive routing

The network-state adaptive baseline updated route cost every 100 ms using normalized delay, delivery loss/ETX, and utilization:with aD=0.40, aQ=0.35, and aU=0.25, subject to aD+aQ+aU=1. Delay, delivery probability, and utilization used the same normalization limits as the proposed method. A route change was installed only when the candidate cost was at least 8% lower than the active-route cost for three consecutive routing updates. Unavailable links were removed before path calculation. A link was treated as degraded when its delivery probability remained below 0.95 or its measured delay exceeded its configured link-delay threshold for three consecutive updates. The baseline used a single FIFO DropTail queue with a maximum capacity of 256 packets per egress interface; it did not use thermal criticality, deterministic gate scheduling, precomputed backup paths, or explicit recovery-confirmation rules as briefly described in Tables 9, 10. Remaining details are included in Table 11.

CpNSA(t)=aDD^p(t)+aQ(1−Q^p(t))+aUU^p(t),(29)

Table 9. Implementation parameters and fairness controls for the four routing configurations.

Table 10. Physical, emulated, and simulated fault-injection protocol.

Table 11. Scenario-based techno-economic and environmental inputs.

4.10.4 TSN-enabled network-state adaptive routing

The TSN-enabled adaptive baseline used exactly the same network-state objective, weights, candidate generation, update interval, and switching rule as the preceding baseline. Its selected flows were mapped to four PCP classes (7, 5, 3, and 0), a 1 ms gate-control cycle, 250 μs safety, 350 μs closed-loop control, 250 μs monitoring, 100 μs background, and 50 μs guard-band windows. IEEE 802.1Qci policing was set to twice the nominal rate for critical streams, and IEEE 802.1Qbu frame preemption was enabled for safety and closed-loop control traffic. This baseline did not use thermal criticality, precomputed backup paths, or the proposed threshold-based recovery-confirmation logic.

4.10.5 Proposed TSDR method

The proposed TSDR method used the same communication metrics, candidate-path generator, controller interface, measurement windows, deterministic queue configuration, and update interval as the strongest baseline. The added components were: (i) the thermal-criticality term Kf(t); (ii) hard delay, reliability, capacity, and availability constraints; (iii) two precomputed feasible backup paths; (iv) fault confirmation after three consecutive violations or missed heartbeats; (v) a 2 s route-hysteresis interval; (vi) recovery verification for five consecutive monitoring windows; and (vii) restoration of the primary path only after ten feasible windows and a cost advantage of at least 10%.

4.11 Experimental scenarios and fault injection

Twelve independent runs are conducted for every routing-architecture and scenario combination.

4.11.1 Normal operation

Each run lasts 6 h, producing 72 h of observations per routing architecture. The tanks begin at 55±0.5∘C with a thermal state of charge of 0.60±0.03. Nominal aggregate thermal demand is 20 kW and average bottleneck-link utilization remains between 30% and 40%.

4.11.2 Peak thermal demand

Each peak-load run lasts 45 min. The aggregate demand command is increased towhich represents three times the nominal system-level thermal demand rather than three times the rated capacity of individual equipment. The physical heat exchangers and pumps remain within their combined safe operating limits.

Q˙peak=3Q˙nom=60kW,(30)

4.11.3 Communication congestion

Background traffic is increased until the principal emulated bottleneck links reach 85%–90% utilization for 10 min intervals. Critical and control traffic rates remain unchanged.

4.11.4 Fault and disturbance conditions

Physical, emulated, and simulated disturbances are applied separately. Seven fault types are evaluated with twelve independent repetitions per routing configuration, resulting in 84 controlled injections for each architecture.

4.12 Performance metrics and measurement sources

For packet k, end-to-end latency is defined asLatency is summarized using the median, P95, and P99 values for each independent run.

Lk=tkreceive−tksend.(31)

Packet-delay variation is defined as the mean absolute difference between consecutive packet delays:Packet Delivery Ratio is calculated asThe deadline-miss ratio isTemperature-tracking error is quantified usingFault-recovery time iswhere stable operation requires compliant latency, PDR, and temperature error for five consecutive monitoring windows.

J=1N−1∑k=2N|Lk−Lk−1|.(32)

PDR=NreceivedNtransmitted×100.(33)

DMR=∑k=1NI(Lk>Dfmax)N×100.(34)

RMSET=1N∑k=1N(Tk−Tkset)2.(35)

Trec=tstable−tfault,(36)

The Energy Utilization Factor is defined over the thermal subsystem aswhere both thermal-energy quantities are calculated from measured inlet/outlet flow and temperature over the run.

EUF=Quseful,deliveredQthermal,supplied,(37)

The system Coefficient of Performance iswhere the denominator includes measured pump, control, and communication electricity. Controller overhead is reported using route-update time, CPU utilization, memory use, and routing-control traffic.

COP=Quseful,deliveredEpump+Econtrol+Enetwork,(38)

4.13 Statistical design

A preliminary power analysis with α=0.05, desired power 1−β=0.80, and a moderate standardized effect size of f=0.25 determined a minimum of ten independently initialized runs per architecture–scenario cell. Twelve runs were used to provide additional allowance for run-level variability. One independently initialized HIL run was the experimental unit; packet-level and sensor-level observations were nested within runs and were not treated as independent replicates.

For continuous run-level outcomes, including P95 latency, jitter, recovery time, temperature RMSE, EUF, and auxiliary power, linear mixed-effects models were fitted usingwhere architecture A, scenario S, and their interaction were fixed effects. Matched seed/run block r was included as a random intercept so that the same initialized condition could be compared across architectures. Runs were nested within scenario in the data structure, while the matched seed identifier was retained across architectures. Evidence domain was not included as a general fixed effect because each primary outcome had a predefined evidence boundary; domain-specific analyses were performed separately when a metric was available from more than one evidence source.

yasr=μ+Aa+Ss+(A×S)as+br+εasr,(39)

PDR and deadline-miss counts were analyzed using beta-binomial generalized linear mixed models with a logit link and the same fixed- and random-effect structure. Model estimation was performed in R version 4.4.2 using the lme4 version 1.1-35.5, lmerTest version 3.1-3, glmmTMB version 1.1.10, and emmeans version 1.10.5 packages. Linear mixed-effects models were estimated using restricted maximum likelihood, while beta-binomial generalized linear mixed models were estimated by maximum likelihood using the Laplace approximation. Degrees of freedom for linear mixed models were calculated using the Satterthwaite approximation. An independent run-level residual covariance structure was used initially, with scenario-specific residual variance terms introduced when residual diagnostics indicated heterogeneous variance.

Missing observations were handled using a predefined run-level completeness rule. A run-level outcome was retained when at least 95% of the required packet or sensor observations were available and the maximum PHY–EMU–SIM synchronization error remained below 0.5 ms. Runs that did not satisfy these requirements were excluded from the corresponding metric analysis before model fitting. No packet- or sensor-level value was substituted after the source-specific stale-data limits described in Section 3.3.3.

Residual-vs.-fitted plots, normal Q–Q plots, and scale-location plots were inspected for continuous outcomes. Variance homogeneity was evaluated using the Brown–Forsythe modification of Levene’s test, and temporal dependence was assessed using residual autocorrelation functions and block-length diagnostics. Beta-binomial models were checked for overdispersion, zero inflation, convergence, and influential architecture–scenario cells. When assumptions were not satisfied, positively skewed outcomes were logarithmically transformed, scenario-specific residual variance structures were fitted, and bootstrap confidence intervals were used when material residual departures remained.

Block-bootstrap confidence intervals with 2,000 resamples were used for temporally autocorrelated packet and thermal summaries. The block length was selected separately for each metric using the first zero crossing of the run-level autocorrelation function, subject to a minimum block duration of 5 s. Primary architecture contrasts compared TSDR with the TSN-enabled network-state adaptive baseline. Pairwise comparisons used Holm-adjusted significance levels. Results are reported with estimated contrasts, standard errors, 95% confidence intervals, adjusted p-values, effect sizes, random-effect variance, residual variance, and intraclass correlation where applicable.

4.14 Reproducibility and data management

Each experimental run is associated with a unique identifier, routing configuration, emulation seed, thermal initial condition, traffic profile, fault schedule, software version, and configuration checksum. Physical, emulated, simulated, and hybrid records are stored in separate timestamped files. Automated validation checks verify timestamp order, unit consistency, missing observations, packet counts, synchronization offsets, and consistency between configuration files and manuscript tables.

The reproducibility package is archived in Zenodo as the Thermal-State-Aware Dynamic Routing HIL Reproducibility Package, release version 1.0.0, under the MIT License for source code and the Creative Commons Attribution 4.0 International License for processed datasets and documentation. The permanent repository DOI is reported in the Data Availability Statement. It contains: (i) the TSDR routing-controller source code; (ii) all three baseline implementations; (iii) NS-3 topology, queue, traffic, TAP-bridge, and fault-injection scripts; (iv) configuration files and random seeds; (v) MATLAB/Simulink and reduced-order CFD parameters required to reproduce the reported states; (vi) processed run-level PHY, EMU, SIM, HYB, and DER datasets; (vii) statistical-analysis scripts; (viii) the scenario-based techno-economic model; and (ix) a README that specifies execution order, software dependencies, expected outputs, and file-level mapping to manuscript tables and figures.

Raw packet captures containing device-level MAC and IP identifiers, vendor-specific PLC project files, proprietary equipment configuration exports, and unprocessed industrial-controller diagnostic logs are subject to controlled access because they contain network-security and vendor-licensed information. Qualified researchers may request access from the corresponding author and the host laboratory, subject to institutional approval, a data-use agreement, and confirmation that the files will be used only for non-commercial research. All processed run-level data required to reproduce the reported numerical comparisons and statistical models are included in the archived package.

5 Scenario-based techno-economic and environmental assessment

The economic and environmental results are projections based on experimental measurements, simulation-derived energy estimates, engineering cost assumptions, and deployment scenarios. They do not represent observed financial performance from an operating industrial facility.

5.1 Assessment boundary and evidence sources

The assessment considers the incremental cost of deploying the proposed thermal-state-aware routing framework relative to a conventional industrial communication and control configuration. The system boundary includes sensors, PLCs, actuators, switches, wireless access points, edge-computing equipment, software integration, installation, commissioning, thermal-storage auxiliaries, maintenance, calibration, data administration, and downtime.

Measured controller and auxiliary-equipment power values are classified as physical evidence (PHY). Annual thermal demand and process-energy savings are obtained from the calibrated thermal model (SIM), while routing and recovery improvements originate from integrated HIL measurements (HYB/DER). Energy tariffs, discount rates, maintenance costs, project lifetime, and carbon-emission factors are treated as scenario assumptions (EST).

5.2 Capital and operational expenditure

Table 12 presents absolute cost estimates in 2026 US dollars. The proposed configuration requires additional sensing, deterministic networking, edge computing, and integration expenditure. These additional costs are partially offset by lower energy consumption, reduced maintenance intervention, and shorter production interruptions.

Table 12. Scenario-based CAPEX and annual OPEX summary for the base deployment scenario.

The itemized cost model is provided in Supplementary Table S24.

5.3 Financial and carbon-reduction models

Annual energy-cost saving is calculated aswhere Ebase,t and Eprop,t denote annual baseline and proposed-system energy consumption, and Cenergy,t is the corresponding energy tariff.

Senergy,t=(Ebase,t−Eprop,t)Cenergy,t,(40)

Net present value is calculated aswhere ΔCAPEX0 is the incremental initial investment, Bt is the annual operating benefit, r is the discount rate, and N is the project lifetime.

NPV=−ΔCAPEX0+∑t=1NBt−ΔOPEXt(1+r)t,(41)

The internal rate of return is the rate satisfyingSimple payback is defined aswhereas discounted payback uses discounted annual cash flows. Return on investment is calculated independently as the ratio of cumulative net benefit to incremental investment and is not treated as equivalent to IRR.

0=−ΔCAPEX0+∑t=1NBt−ΔOPEXt(1+IRR)t.(42)

PBsimple=ΔCAPEX0Bannual,(43)

Projected annual carbon saving iswhere EFgrid is the grid-emission factor.

CO2,saved=(Ebase−Eproposed)EFgrid,(44)

The projected 112 MWh annual energy saving comprises approximately 103 MWh from improved thermal charge/discharge coordination and reduced thermal losses, and 9 MWh from lower pump, control, and communication energy. Using EFgrid=0.61tCO2e/MWh, the base scenario givesUnder the base assumptions, the incremental investment has an NPV of approximately USD 104,000, an IRR of 41.8%, a simple payback of 2.32 years, and a discounted payback of 2.68 years.

CO2,saved=112×0.61=68.3tCO2e/year.(45)

5.4 Economic uncertainty and sensitivity analysis

Sensitivity analysis evaluates conservative, base, and optimistic deployment scenarios by varying incremental CAPEX, annual energy saving, electricity price, discount rate, project lifetime, maintenance saving, downtime cost, and the carbon-emission factor.

The sensitivity results indicate that energy saving, downtime valuation, and incremental integration cost have the largest influence on NPV and payback. The conservative case remains economically positive but has a substantially longer discounted-payback period.

5.5 Potential application settings

Potential application settings include retrofit thermal-storage facilities, industrial energy-management systems, district-energy demonstrators, and renewable-energy-integrated storage plants. Progression beyond the present prototype requires a field pilot, multi-vendor interoperability testing, long-duration operation, safety certification, cybersecurity assessment, and larger-scale physical validation as mentioned in Tables 13, 14.

Table 13. Scenario-based financial and environmental sensitivity results.

Table 14. Hybrid HIL verification and data-quality results.

6 Experimental results

6.1 Hybrid HIL verification and data quality

A total of 480 independently initialized runs were completed across the four routing architectures and ten operating or fault conditions. All planned runs satisfied the predefined data-quality criteria. The final dataset contained more than 1.2×107 packet events and 8.6×106 synchronized thermal observations; these packet- and sensor-level records were not treated as independent experimental replicates.

Overall data completeness was 99.93%. The median physical-clock synchronization error was 18μs, with a maximum of 47μs. Median packet-bridge delay between the physical gateway and NS-3 was 0.42 ms, while its P95 value was 0.71 ms. Real-time thermal-step compliance was 99.98%, and the NS-3 wall-clock execution ratio was 0.9997±0.0003, further details are mentioned in Table 15.

Table 15. Physical-subsystem performance under normal and disturbance conditions.

Digital-twin validation produced a temperature RMSE of 0.42∘C, a flow MAE of 0.28Lmin−1, and a pressure MAE of 2.1 kPa. These values were within the predefined model-validation limits.

6.2 Physical-subsystem results

Physical measurements showed improved thermal tracking and actuator response under the proposed routing method. Relative to the TSN-enabled network-state adaptive baseline, the proposed method reduced physical temperature RMSE from 0.81 to 0.62∘C, reduced flow MAE from 0.57 to 0.46Lmin−1, and shortened valve settling time from 1.41 to 1.18 s.

The measured power boundary includes the circulation pumps, PLCs, valves, physical network equipment, edge controller, and data-acquisition system. It does not include simulated thermal loads.

6.3 Integrated hybrid communication and energy performance

Table 16 summarizes the principal results aggregated across normal, peak-load, congestion, and fault windows. Values are run-level estimates with 95% confidence intervals. Improvements are calculated relative to the strongest baseline, namely TSN-enabled network-state adaptive routing.

Table 16. Comparative physical and hybrid performance across routing architectures.

The proposed method reduced P95 latency by 34.7%, jitter by 25.0%, and recovery time by 52.5% relative to the TSN-enabled network-state adaptive baseline. PDR increased by 1.8 percentage points. The improvement in EUF reflects better coordination between communication priority and thermal charge/discharge control. These effects apply to the implemented hybrid HIL topology, traffic profiles, queue configuration, thermal initial conditions, and fault schedules; they do not establish general industrial deployment performance.

6.4 Statistical analysis

The statistical unit was one independently initialized run. Each architecture–condition cell contained 12 runs, with a 10 min warm-up interval followed by the scenario-specific measurement period. Routing tables, controller states, physical thermal initial conditions, and emulation seeds were reset before each run. Experimental order was randomized within matched seed blocks. The model structure and diagnostics followed Section 4.13.

Table 17 reports the prespecified TSDR-vs.-TSN-adaptive contrasts. Confidence intervals for latency, jitter, recovery time, temperature RMSE, and EUF are model-based intervals verified using run-level block bootstrap; PDR and deadline-miss intervals are reported on the odds-ratio scale from the beta-binomial models.

Table 17. Prespecified TSDR-vs.-TSN-adaptive statistical contrasts under the evaluated HIL conditions.

The architecture-by-scenario interaction was statistically significant for P95 latency (F18,396=6.31, p<0.001), jitter (F18,396=4.86, p<0.001), recovery time (F15,330=4.42, p<0.001), temperature RMSE (F18,396=3.74, p<0.001), and EUF (F18,396=2.69, p<0.001). Likelihood-ratio tests also identified architecture-by-scenario interactions for PDR (χ182=39.8, p=0.0022) and deadline misses (χ182=42.6, p=0.0009). For P95 latency, the matched-run random-intercept variance was 22.7ms2, residual variance was 58.4ms2, and the intraclass correlation was 0.28. Corresponding variance components for recovery time and temperature RMSE are reported in Supplementary Table S25. Residual-vs.-fitted and Q–Q diagnostics showed no material departures from linearity, residual normality, or variance homogeneity after logarithmic transformation of recovery time; the beta-binomial models showed acceptable dispersion ratios of 1.07 for PDR and 1.11 for deadline misses, with no evidence of material zero inflation. The supplementary diagnostic figure provides the model checks for P95 latency, recovery time, and temperature RMSE.

The statistical findings identify differences under the evaluated HIL conditions. They do not establish equivalent effects in a full industrial deployment, a larger physical network, or an untested operating environment.

6.5 Scenario-based performance

Table 18 presents proposed-method performance under the principal experimental scenarios.

Table 18. Scenario-specific results for the proposed routing method.

6.6 Ablation analysis

Ablation experiments isolate the effects of network adaptation, thermal-state awareness, and self-healing.

Thermal-state awareness produced the largest reduction in latency and temperature error, whereas the explicit self-healing mechanism produced the largest reduction in fault-recovery time. The complete model achieved the best combined communication and thermal performance.

6.7 Weight-sensitivity analysis

Each thermal-criticality and route-cost weight was varied independently by −20% and +20%, followed by renormalization of the corresponding weight vector. All other parameters, seeds, traffic profiles, and fault schedules were held constant. Table 20 reports the resulting run-level outcomes.

Across the verified perturbation set, the TSDR ranking against the strongest TSN-enabled network-state adaptive baseline remained unchanged for every evaluated outcome. The largest observed degradation relative to the calibrated configuration was a 6.4% increase in P95 latency, a 0.6-percentage-point reduction in PDR, an 8.6% increase in recovery time, a 5.9% increase in temperature RMSE, and a 12.4% increase in route-switch frequency. The sensitivity analysis is interpreted only as local robustness around the calibrated deterministic weights; it does not establish optimality outside the evaluated parameter range.

6.8 Emulation-based scalability

Scalability beyond the integrated 160-node topology was evaluated as EMU evidence through network emulation while retaining physical sensing, PLC, gateway, timestamping, and actuation only at the HIL boundaries. Results above 160 logical nodes therefore characterize controller behavior in larger emulated topologies and do not represent physically deployed scalability.

The worst-case route-update time remained below the 100 ms routing interval at 640 logical nodes. However, CPU utilization and control overhead increased substantially at the largest emulated scale. This EMU result does not establish timing, reliability, interoperability, or fault-recovery performance for a physically deployed 640-node industrial network as mentioned in Tables 19, 21.

Table 19. Ablation analysis of the principal routing components.

Table 20. Sensitivity of principal outcomes to ±20% single-weight perturbations.

Each entry reports the −20%/+20% result.

Table 21. Emulation-based scalability of the routing controller.

7 Discussion

7.1 Interpretation and HIL-level implications

The physical and integrated hybrid results indicate that communication-aware prioritization influenced both network performance and thermal-control quality. Thermally critical flows were assigned lower-cost paths and higher queue priorities, thereby reducing delayed control actions and temperature-tracking error. Congested links were penalized through the utilization component of the routing objective, while unavailable links were excluded through the path-availability constraint. Precomputed backup paths reduced the interval between fault confirmation and route activation, route hysteresis limited unnecessary switching between paths with similar costs, and deterministic queue scheduling reduced packet-delay variation. The hybrid results further showed that combining thermal criticality with network-state adaptation reduced P95 latency, jitter, deadline misses, and fault-recovery time relative to network-only routing. Emulation-derived results supported repeatable fault testing and scalability analysis, while simulation-derived thermal states enabled controlled process disturbances and the representation of additional virtual assets. Physical, emulated, simulated, and integrated hybrid evidence was analyzed separately to avoid interpreting emulation- or simulation-derived outcomes as direct physical measurements. Within the evaluated HIL configuration, the principal technical contribution is the formal inclusion of thermal-process criticality in a constrained communication-route objective. The route-cost model in Equation 11 combines normalized delay, jitter, delivery probability, congestion, energy cost, fault risk, and thermal urgency. A second contribution is the self-healing mechanism, which defines fault-confirmation thresholds, precomputed backup paths, switching rules, route hysteresis, recovery conditions, and return-to-primary criteria. The ablation analysis demonstrates that thermal-state awareness primarily improves latency and temperature tracking, whereas self-healing primarily reduces fault-recovery time.

A further contribution is the integration of physical measurements (PHY), network-emulation outputs (EMU), simulation-derived thermal states (SIM), and integrated hybrid measurements (HYB) within one timestamp-aligned control loop. The hybrid HIL methodology also provides a reproducible evaluation framework that distinguishes measured observations, emulation-derived results, simulation outputs, derived metrics, and scenario-based economic projections. The proposed method nevertheless introduces additional sensing, synchronization, computation, and integration requirements. Accurate thermal criticality depends on calibrated temperature, pressure, flow, and state-of-charge information; therefore, physical sensor errors or digital-twin parameter uncertainty can affect traffic prioritization. The routing controller also introduces greater control overhead than static routing. Although the evaluated controller maintained the 100 ms update deadline, computation time and memory use increased with logical node count. Route hysteresis reduced path oscillation but may delay restoration of a recovered primary route.

Implementation additionally requires deterministic queue configuration, synchronized timestamps, gateway integration, and compatible industrial interfaces. These requirements increase initial CAPEX and may require vendor-specific engineering. The environmental relevance of the framework therefore depends on measurable thermal-energy savings, operating duration, electricity mix, and site-specific economic conditions rather than communication improvement alone.

7.2 Limitations and transferability

The physical subsystem consisted of selected thermal tanks, sensors, PLCs, actuators, gateways, and industrial network devices. Additional communication nodes, links, traffic sources, and network faults were represented through real-time network emulation, while additional thermal assets and selected process disturbances were represented using synchronized simulation models. The findings therefore characterize a hybrid physical–emulation HIL environment rather than a complete industrial field deployment. The physical communication boundary contained 14 network endpoints, the integrated hybrid topology contained 160 physical and emulated nodes, and the scalability analysis extended to 640 logical nodes through emulation. These physical, integrated, and emulation-only scales should be interpreted separately. Emulation-based scalability results demonstrate controller behavior under larger logical topologies but do not replace large-scale physical validation. The evaluation used 12 independent runs for each architecture, scenario combination. Although this design supported run-level statistical analysis, packet-level and sensor-level observations remained temporally autocorrelated and were therefore summarized per run or analyzed using block-bootstrap methods. The controlled laboratory environment represented predefined traffic, thermal-demand, and fault conditions. Electromagnetic interference, multi-vendor protocol behavior, seasonal thermal variation, long-term equipment degradation, operator-induced disturbances, and real-world traffic diversity were more limited than in a continuously operating industrial facility.

Digital-twin outputs depend on calibrated thermal parameters and may be affected by model mismatch outside the evaluated temperature, flow, pressure, and demand ranges. Fault duration and combined-fault complexity were also constrained by the experimental safety protocol. Performance may therefore change in multi-site industrial systems with different radio environments, traffic policies, ownership structures, and control hierarchies. Application to steam, compressed-air, district-heating, or high-temperature manufacturing systems would require domain-specific state variables, safety constraints, thermal models, and route-priority thresholds. Economic and environmental outcomes also depend on regional electricity tariffs, labor costs, project lifetime, carbon intensity, annual operating hours, and downtime valuation. The financial and carbon results are scenario-based projections and should not be interpreted as realized field outcomes. The results are insufficient to establish industrial deployment readiness, safety certification, long-term reliability, or plant-level economic benefit.

7.3 Future research and field validation

Future work should validate the deterministic TSDR method through long-duration, multi-vendor industrial pilots with larger physical endpoint counts, broader radio and traffic conditions, cybersecurity testing, operator-induced disturbances, and safety-certification requirements. Additional studies should examine predictive fault detection, digital-twin uncertainty, secure OPC UA/MQTT operation, alternative constrained optimization methods, and learning-based weight adaptation. Any learning-based extension should be evaluated separately from the present deterministic method using explicit training, calibration, and testing partitions, while preserving hard safety and communication constraints. Reproducible benchmark configurations, open implementations, and transparent field-level evidence should remain priorities before claims of broader industrial readiness are made.

8 Conclusion

This study developed a deterministic Thermal-State-Aware Dynamic Routing (TSDR) method for industrial thermal-storage communication. The method combines physical-process criticality with delay, jitter, packet-delivery probability, congestion, link availability, communication-energy cost, and fault risk within a constrained route-selection objective. Its self-healing mechanism incorporates heartbeat-based fault detection, precomputed backup paths, route hysteresis, and recovery-confirmation criteria, while deterministic queue scheduling provides differentiated service for safety, control, monitoring, and background traffic. The method was evaluated through a hybrid physical–emulation HIL framework that separated physical measurements, network-emulation outputs, simulation-derived thermal states, integrated hybrid measurements, derived indicators, and scenario-based estimates.

Under the evaluated HIL conditions, relative to the TSN-enabled network-state adaptive baseline, TSDR reduced P95 latency by 34.7%, jitter by 25.0%, and fault-recovery time by 52.5%. It increased PDR by 1.8 percentage points, reduced deadline misses by 3.9 percentage points, improved EUF from 0.84 to 0.95, and lowered measured auxiliary power from 4.05 to 3.68 kW. Physical temperature-tracking RMSE decreased from 0.81 to 0.62∘C. The statistical analysis identified significant differences under the implemented topology, traffic, queue, thermal, and fault configurations, with moderate-to-large effect sizes for the principal outcomes. The ablation analysis indicated that thermal-state awareness primarily affected latency and thermal tracking, whereas the explicit recovery mechanism produced the largest reduction in recovery time.

Scalability beyond the integrated 160-node topology was examined through network emulation rather than physical deployment. The controller maintained its 100 ms update deadline at 640 logical nodes in the EMU tests, although CPU utilization, memory consumption, and routing overhead increased with logical network size. The scenario-based economic model projected an NPV of approximately USD 104,000, an IRR of 41.8%, a simple payback of 2.32 years, a discounted payback of 2.68 years, and annual carbon savings of 68.3tCO2e under the base assumptions. These values are EST projections derived from HIL-informed scale-up, simulation outputs, and documented economic assumptions; they are not measured plant-level benefits.

Overall, the findings indicate, under the evaluated conditions, that coupling thermal-process urgency with communication-state adaptation can improve communication resilience and thermal-control tracking in the reported hybrid HIL configuration. The study does not establish full industrial deployment readiness, long-term field reliability, safety certification, large-scale physical scalability, or realized economic and environmental performance.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fhumd.2026.1914527/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Serhan S. Alshammari, University of Hail, Saudi Arabia

Reviewed by: Djamel Ghernaout, University of Hail, Saudi Arabia

Anwar Shah, National University of Computer and Emerging Sciences, Pakistan

Christos Tsallis, University of West Attica, Greece

Abbreviations CAPEX, Capital Expenditure; CI, Confidence Interval; COP, Coefficient of Performance; DER, Derived Metric; EMU, Network-Emulation Evidence; EST, Scenario-Based Estimate; ETX, Expected Transmission Count; EUF, Energy Utilization Factor; FRT, Fault Recovery Time; HIL, Hardware-in-the-Loop; HYB, Integrated Hybrid Evidence; IIoT, Industrial Internet of Things; NPV, Net Present Value; OPEX, Operational Expenditure; OPC UA, Open Platform Communications Unified Architecture; PDR, Packet Delivery Ratio; PHY, Physical Evidence; PLC, Programmable Logic Controller; SDN, Software-Defined Networking; SIM, Simulation-Derived Evidence; TSN, Time-Sensitive Networking; TSDR, Thermal-State-Aware Dynamic Routing.