PaperlayerView exact revision on arXiv

Source-linked AI summary

A Hybrid Edge Cloud Digital Twin for Welfare-Constrained Control in Poultry Production

Suresh Neethirajan

arXiv:2608.20367v1eess.SYcs.AI

TL;DR

Poultry climate control remains largely heuristic despite biologically variable environments and practical connectivity limits. The paper introduces an edge–cloud digital twin combining physical and learned models with state estimation and welfare-constrained MPC. Simulation results show lower prediction error, fewer ammonia violations, and substantially reduced communication requirements, while limitations remain in validation scope and real-world deployment.

  • Problem

    Commercial poultry environmental control remains largely heuristic despite biological variability, welfare constraints, sparse data, and connectivity-limited facilities.

  • Method

    An edge–cloud digital twin integrates grey-box thermodynamic and mass-balance modeling, learned residuals, edge state estimation, and welfare-constrained model predictive control.

  • Results

    Temperature prediction error fell from 1.8 °C to 0.4 °C, ammonia violations decreased by 90%, and communication requirements fell approximately 30× through edge-first processing.

  • Takeaways & Limitations

    The results suggest that physically grounded digital twins can support reliable closed-loop, welfare-aware control in biologically complex environments.

  • Takeaways & Limitations

    Validation used only two facility typologies and physics-based numerical simulation; substantially different housing and commercial deployment require further verification.

Abstract

from arXiv · show

Poultry production operates under tightly coupled environmental and biological dynamics, yet commercial climate control remains largely heuristic, limiting welfare assurance and operational efficiency. We introduce an edge-cloud digital twin framework for real-time, welfare-constrained environmental control in poultry facilities. The framework integrates distributed sensing, on-device state estimation, a hybrid physics-data model, and model predictive control to enable anticipatory and adaptive management under practical farm constraints. A grey-box thermodynamic and mass-balance formulation is augmented with a learned residual that captures unmodeled biological variability, including activity-dependent metabolic heat. This hybrid model is embedded within a state-space representation for real-time estimation and control at the edge, while cloud coordination supports cross-farm learning and long-horizon optimization. Bandwidth-aware processing and asynchronous synchronization enable deployment in connectivity-limited environments. Evaluation in a high-fidelity broiler production testbed demonstrates substantial gains over rule-based control and physics-only modeling. Temperature prediction error is reduced from 1.8 degrees Celsius to 0.4 degrees Celsius, ammonia constraint violations decrease by 90 percent, and communication requirements are lowered approximately 30-fold through edge-first processing. A Domain Transfer Score of 0.92 further indicates strong robustness across facility conditions. These results show that physically grounded digital twins, coupled with real-time control, enable scalable and welfare-aware management of biological production systems.

I. INTRODUCTION

Poultry facilities face stochastic biological dynamics, welfare-sensitive environmental risks, and connectivity constraints that make heuristic, reactive control inadequate. The proposed edge–cloud digital twin combines physical modeling, hybrid learning, and distributed computation for real-time evaluation and welfare-constrained management.

  • 130 billion birds are produced globally each year, while commercial poultry control remains largely heuristic, manually supervised, and reactive.
  • Biological variability arises from animal behavior, microbial processes, weather, sparse ground-truth data, and heterogeneous legacy infrastructure.
  • Physics-based simulations are computationally intensive, whereas machine-learning approaches often lack interpretability, require large labeled datasets, and generalize poorly across facilities.
  • The framework embeds conservation-based modeling, hybrid physics–data learning, state-space control, and hard welfare constraints within an edge–cloud architecture.
  • A high-fidelity broiler testbed evaluates prediction accuracy, control performance, and deployment feasibility before real-world implementation.

A. Digital Twins and Industrial Cyber-Physical Systems

Digital-twin and cyber-physical methods provide architectural and control foundations, but livestock applications remain weakly integrated across welfare constraints, distributed execution, and cross-farm transfer. This work positions a unified framework to address that gap.

  • A. Digital Twins and Industrial Cyber-Physical Systems: Industrial digital twins represent physical assets through virtual counterparts that receive sensor updates and support predictive and what-if analyses.
  • A. Digital Twins and Industrial Cyber-Physical Systems: Industry 4.0 reference architectures and Asset Administration Shell semantics provide foundations for interoperable digital-twin deployment.
  • B. Environmental Control and HVAC Systems: Grey-box resistance–capacitance models, mass balances, state-space formulations, and model predictive control offer computationally efficient methods for dynamic thermal and mass transport systems.
  • C. Livestock Environmental Modeling and Control: Livestock models capture bioenergetic heat and moisture production, while ventilation and welfare metrics address ammonia, carbon dioxide, and physiological stress.
  • C. Livestock Environmental Modeling and Control: Most commercial poultry systems remain rule-based, while sensing and machine-learning platforms generally lack formal control-theoretic integration.
  • D. Gap Statement and Novel Positioning: The remaining gap is a unified architecture combining welfare-constrained optimization, edge–cloud execution, semantic interoperability, and cross-farm generalization.
  • D. Gap Statement and Novel Positioning: Domain Transfer Score measures prediction-accuracy retention when a calibrated digital twin is applied to a new facility without retraining.
  • D. Gap Statement and Novel Positioning: The proposed approach integrates state-space modeling, model predictive control, and hybrid physics–data learning for biologically nonstationary systems.

III. PROBLEM FORMULATION AND SYSTEM OVERVIEW

The poultry house is modeled as a cyber-physical system linking biological processes, distributed sensing, edge computation, cloud coordination, and welfare-constrained control. Its layered architecture assigns fast operational functions to the edge and longer-horizon analytics to the cloud.

  • A. Poultry Production as an Industrial Cyber-Physical System: A representative broiler house combines ventilation, heating, inlet dampers, and biological thermal and moisture loads that vary with flock conditions.
  • A. Poultry Production as an Industrial Cyber-Physical System: Distributed sensors measure temperature, humidity, carbon dioxide, and ammonia, while vision sensing tracks flock distribution and activity.
  • A. Poultry Production as an Industrial Cyber-Physical System: The edge gateway performs local state estimation, control execution, and feature extraction, while the cloud supports planning, retraining, cross-farm analytics, and logging.
  • A. Poultry Production as an Industrial Cyber-Physical System: Welfare is enforced as a hard constraint while energy, cost, and equipment limits are optimized within the control problem.
  • B. High-Level System Architecture: The architecture spans physical sensing, edge computing, digital-twin execution, supervisory control, and analytics across distinct timescales.

IV. MATHEMATICAL FORMULATION OF THE PROPOSED EDGE-CLOUD DIGITAL TWIN FRAMEWORK

The framework combines grey-box thermal and mass-balance models with state-space estimation, residual learning, and welfare-constrained MPC. This design preserves physical consistency while supporting efficient edge execution and adaptive prediction.

  • A. Thermal Dynamics and Environmental Modeling: The barn thermal model uses a lumped resistance–capacitance network derived from energy-balance principles.
  • A. Thermal Dynamics and Environmental Modeling: Metabolic heat is modeled as a time-varying internal input that depends on bird mass, activity level, and age, alongside ventilation, heating, solar, and loss terms.
  • A. Thermal Dynamics and Environmental Modeling: The thermal formulation preserves thermodynamic consistency while remaining computationally efficient for real-time edge execution.
  • B. Air Quality and Mass Transport Dynamics: Mass-balance dynamics model humidity, carbon dioxide, and ammonia concentrations using flow rates, inlet concentration, and biologically driven generation.
  • B. Air Quality and Mass Transport Dynamics: Explicit mass-balance prediction enables control of welfare-critical air-quality variables rather than using temperature as a proxy.
  • C. Cyber-Physical System State-Space Representation: Coupled thermal and mass-balance dynamics are expressed in state-space form for integration with estimation and control algorithms.
  • C. Cyber-Physical System State-Space Representation: System matrices are physically derived or calibrated using recursive least squares, enabling Kalman filtering and model predictive control.
  • D. Supervisory Control and Optimization: MPC optimizes welfare deviations, control effort, energy, and cost over a finite horizon while enforcing actuator and welfare limits.

V. SYSTEM ARCHITECTURE OF THE PROPOSED EDGE– CLOUD DIGITAL TWIN

The proposed digital twin uses a layered edge–cloud architecture that combines distributed sensing, local state estimation, hybrid modeling, supervisory welfare-constrained control, and cloud coordination. Edge execution supports real-time responsiveness and bandwidth reduction, while cloud services provide analytics, governance, and longer-horizon optimization.

  • Architecture: The architecture comprises five layers spanning physical sensors and actuators, edge computing, cyber-physical modeling, supervisory control, and cloud analytics.These layers connect facility instrumentation to digital-twin computation, control optimization, and cloud governance.
  • Edge Computing: The edge layer normalizes heterogeneous sensor and legacy-equipment signals, extracts features locally, and reduces raw-data transmission to the cloud.High-frequency signals are summarized locally, video is converted to detected events, and asynchronous data is handled separately.
  • Digital Twin Core: The distributed digital-twin core assigns real-time control authority to the edge twin while the cloud twin supports analytics and long-horizon planning.Soft synchronization permits limited divergence under bandwidth constraints, whereas hard synchronization can transfer full snapshots or override setpoints for safety-critical scenarios.
  • Supervisory Control: The edge MPC executes every 5 minutes with a 30–60 minute prediction horizon, optimizing welfare-setpoint tracking subject to actuator constraints.The loop issues setpoints for ventilation fans, heaters, and dampers while the welfare-constrained MPC diagram includes state estimation from noisy measurements.
  • Cloud and Interoperability: The cloud layer provides dashboards, KPI tracking, anomaly detection, model retraining, and strategic oversight across connected facilities.The framework also uses AAS-based submodels to represent facility identity, technical data, operational data, welfare metrics, and digital-twin configuration.

VI. SIMULATION-BASED VALIDATION

Validation uses a high-fidelity numerical broiler-house testbed and a three-phase simulation study. The design evaluates hybrid-model accuracy, edge–cloud bandwidth coordination, and welfare-constrained MPC against physics-only and rule-based baselines.

  • Testbed: The testbed represents a 120 meter by 15-meter commercial broiler house containing approximately 25,000 Ross 308 birds over a 42-day flock cycle.The simulation models representative facility, flock, sensing, and equipment conditions before real-world deployment.
  • Validation Design: The validation spans calibration, bandwidth optimization, and full-cycle welfare-constrained MPC evaluation across approximately 12 weeks of simulated operation.The phases isolate prediction accuracy, communication efficiency, and control performance.
  • Edge–Cloud Evaluation: The edge-first evaluation transmits aggregated statistics or detected anomalies instead of raw high-frequency sensor and video streams.This phase tests bandwidth reduction under cloud-connected and connectivity-limited scenarios.
  • Control Evaluation: Full-cycle MPC is compared with a conventional rule-based controller under seasonal temperature, humidity, and flock-growth variations.The comparison tests maintenance of the thermoneutral zone under representative broiler-house operating conditions.
  • Model Validation: Hybrid-model predictions are trained on synthetic flock-cycle data and evaluated on unseen simulation data using RMSE and MAPE for temperature, CO2, and ammonia.The residual learner combines physics-state estimates, weather, temporal features, and lagged errors to model unmodeled biological heat patterns.

VII. RESULTS

The simulated results compare the proposed edge–cloud digital twin with existing approaches and a rule-based controller across prediction, welfare, and transfer metrics. The framework combines capabilities within one architecture and reports improved accuracy, time in range, and cross-facility generalization.

  • Comparison: The proposed framework combines modeling, control, and edge–cloud capabilities within a unified architecture compared with approaches addressing individual poultry-modeling aspects.Table I positions the framework qualitatively against existing poultry modeling approaches.
  • Generalization: 0.92 Domain Transfer Score indicates that the calibrated digital twin retains 92% of its prediction accuracy when transferred to new facilities without retraining.The score compares transfer error with a facility-agnostic baseline using nominal physical parameters.
  • Prediction Accuracy: 1.8 to 0.4 degrees Celsius Temperature Prediction RMSE represents a 77% reduction relative to the baseline.The metric compares the proposed digital twin controller with the baseline values reported in Table II.
  • Prediction Accuracy: 24% to 6% Ammonia Prediction MAPE represents a 75% reduction relative to the baseline.This result is reported among the simulation performance metrics.
  • Welfare Control: 82% to 96% Time-in-Range (TNZ) represents a 14% increase under the proposed controller.The metric reports the share of operation within the thermoneutral zone.

ne (RBC)

The reported metrics show lower energy use, sharply reduced bandwidth demand, and strong transfer robustness for the proposed framework.

  • 15% reduction in Total Energy Use, from 4,200 to 3,570 kilowatt hours per cycle.
  • Bandwidth Usage falls from 5,400 to 180 megabytes per day.
  • Domain Transfer Score reaches 0.92, indicating High Robustness.The accompanying note defines 1.0 as perfect transfer to a new farm without retraining.
  • RMSE is defined as Root Mean Square Error, with lower values indicating better performance.

VIII. DISCUSSION

The framework improves prediction, welfare-constrained control, and connectivity efficiency, while remaining subject to calibration, transfer, and validation limitations. Its standards alignment also supports prospective commercial integration and welfare reporting.

  • Interpretation of Results: 0.4 degrees Celsius temperature prediction RMSE versus 1.8 degrees Celsius for the pure physics baseline.The neural network residual learner captures nonlinear metabolic heat patterns that vary with flock activity and distribution.
  • Interpretation of Results: Up to 90% reduction in ammonia violations demonstrates the welfare-constrained MPC’s ability to prioritize animal health over short-term energy savings.
  • Robustness and Edge-Cloud Efficacy: 100% control continuity was maintained despite intermittent cloud access under rural connectivity constraints.The edge-first architecture and AAS-based semantic interoperability supported this continuity.
  • Robustness and Edge-Cloud Efficacy: Bandwidth reduction from 5.4 gigabytes per day to 180 megabytes per day supports deployment in infrastructure-poor agricultural regions without sacrificing digital-twin state fidelity.
  • Limitations: The hybrid residual model requires periodic ground-truth calibration to prevent drift over multiple flock cycles.The stated example is manual litter-moisture checking.
  • Limitations: Validation covered only two facility typologies, and substantially different housing styles may require re-identification of the state-space matrices.The linear state-space formulation may also reduce accuracy in barn corners or obstructed zones by simplifying complex fluid dynamics.
  • Future Work: Real-world commercial-farm deployment remains necessary to verify performance claims because validation used physics-based numerical simulation.The stated deployment considerations include biological variability, sensor drift, and equipment failures.
  • Regulatory and Commercial Pathway: Alignment with ISO 23247 and IEC 63278 supports standardized data exchange, welfare-certification reporting, supply-chain traceability, and integration with existing systems.

IX. CONCLUSIONS

The proposed edge–cloud digital twin combines physically grounded modeling, state estimation, hybrid learning, and welfare-constrained control for real-time poultry environmental management. Simulation results show improved prediction, welfare compliance, communication efficiency, and cross-facility robustness, while real-world validation remains necessary.

  • IX. CONCLUSIONS: The system integrates thermodynamic and mass-balance modeling, state-space estimation, hybrid physics–data learning, and welfare-constrained model predictive control.This integration supports real-time environmental management that is computationally efficient and operationally relevant.
  • IX. CONCLUSIONS: Temperature prediction error is reduced from 1.8 °C to 0.4 °C, ammonia constraint violations decrease by 90%, and communication requirements fall approximately 30× through edge-first processing.The Domain Transfer Score of 0.92 indicates strong robustness across facilities.
  • IX. CONCLUSIONS: Enforcing welfare as a hard constraint and maintaining physical consistency within learning improves robustness and practical deployability.The framework integrates modeling, control, and distributed computation within a unified cyber-physical architecture.
  • IX. CONCLUSIONS: Validation is currently limited to a high-fidelity numerical testbed and should be extended to real-world conditions with sensor noise, equipment variability, and broader facility diversity.Future directions include federated learning, adaptive control, and multi-species deployment.
  • IX. CONCLUSIONS: The results support scalable, welfare-aware environmental control and movement beyond monitoring toward actionable decision-making in biologically complex production systems.The conclusion presents edge–cloud digital twins as a viable pathway for agricultural cyber-physical automation.
Loading 2608.20367v1…