arxiv: 2604.13396 · v1 · submitted 2026-04-15 · 📡 eess.SY · cs.SY

Recognition: unknown

HierFedCEA: Hierarchical Federated Edge Learning for Privacy-Preserving Climate Control Optimization Across Heterogeneous Controlled Environment Agriculture Facilities

Andrii Vakhnovskyi

Authors on Pith no claims yet

Pith reviewed 2026-05-10 13:21 UTC · model grok-4.3

classification 📡 eess.SY cs.SY

keywords federated learningcontrolled environment agricultureclimate controldifferential privacyhierarchical modelsPID auto-tuningHVAC optimizationedge learning

0 comments

The pith

HierFedCEA achieves 94 percent of centralized performance for privacy-preserving climate control optimization across heterogeneous CEA facilities.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents HierFedCEA, a hierarchical federated learning framework that lets controlled environment agriculture facilities collaborate on HVAC optimization without exchanging raw operational data containing sensitive grow recipes. It decomposes the neural PID auto-tuning model into three tiers aligned with physical structure: a global physics layer, a crop-cluster layer for cultivar-specific mappings, and a local personalization layer for equipment dynamics. Tier-specific differential privacy is applied to the compact 36-parameter model, yielding privacy essentially at no extra cost. Simulations calibrated on seven-plus years of data from over thirty facilities across eight U.S. climate zones show 94 percent of centralized performance, total communication below one megabyte, and excess risk under 0.15 percent. This matters because current data-sharing refusals block the 30-38 percent energy reductions and faster commissioning that cross-facility knowledge transfer could deliver.

Core claim

HierFedCEA decomposes the neural network PID auto-tuning model into three tiers aligned with the physical structure of the control problem—a global physics tier capturing universal thermodynamic relationships, a crop-cluster tier encoding cultivar-specific VPD-to-gain mappings, and a local personalization tier adapting to facility-specific equipment dynamics—then applies tier-specific differential privacy budgets to enable privacy-preserving cross-facility optimization while achieving 94 percent of centralized training performance at under 1 MB total communication cost and less than 0.15 percent excess risk.

What carries the argument

The three-tier hierarchical decomposition of the compact 36-parameter neural PID auto-tuning model, with tier-specific differential privacy budgets aligned to global physics, crop-cluster, and local equipment layers.

Load-bearing premise

Simulation experiments calibrated from historical production data across multiple facilities will accurately predict real-world performance and privacy-utility trade-offs when deployed on live heterogeneous equipment.

What would settle it

A live deployment on operational facilities that measures actual energy savings, model convergence, and any observed privacy leakage against the simulated 94 percent performance and 0.15 percent excess risk figures.

Figures

Figures reproduced from arXiv: 2604.13396 by Andrii Vakhnovskyi.

read the original abstract

Cross-facility knowledge transfer in Controlled Environment Agriculture (CEA) can reduce HVAC energy consumption by 30-38% and accelerate new facility commissioning from months to days. However, facility operators refuse to share raw operational data because it encodes commercially sensitive grow recipes. We present HierFedCEA, a hierarchical federated learning framework that enables privacy-preserving climate control optimization across heterogeneous CEA facilities. HierFedCEA decomposes the neural network PID auto-tuning model into three tiers aligned with the physical structure of the control problem: (1) a global physics tier capturing universal thermodynamic relationships; (2) a crop-cluster tier encoding cultivar-specific VPD-to-gain mappings; and (3) a local personalization tier adapting to facility-specific equipment dynamics. The framework applies tier-specific differential privacy budgets and leverages the extreme compactness of the 36-parameter PID model to achieve privacy essentially for free (excess risk < 0.15%). Simulation experiments calibrated from 7+ years of production deployment across 30+ commercial facilities in 8 U.S. climate zones demonstrate that HierFedCEA achieves 94% of centralized training performance while reducing total communication cost to under 1 MB. To the best of our knowledge, this is the first federated learning framework designed for CEA climate control.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

HierFedCEA introduces a physics-aligned three-tier FL setup for CEA climate control that keeps communication tiny via a compact PID model, but all performance numbers come from unvalidated simulations.

read the letter

The main point is that this paper puts forward a hierarchical federated learning framework for optimizing climate control across different CEA facilities without sharing raw data. It splits the model into a global physics tier, a crop-cluster tier, and a local personalization tier, which lines up with how these systems are built in practice. That structure plus the 36-parameter PID compactness lets them add tier-specific differential privacy with almost no performance hit and total communication under 1 MB while reaching 94 percent of centralized results in their tests. Those are the concrete engineering moves that stand out as new for this domain.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces HierFedCEA, a hierarchical federated learning framework for privacy-preserving climate control optimization across heterogeneous CEA facilities. It decomposes a neural-network PID auto-tuner into three tiers (global physics, crop-cluster VPD-to-gain mappings, and local equipment personalization), applies tier-specific differential privacy budgets, and leverages the compactness of the 36-parameter PID model. Simulation experiments calibrated on 7+ years of production data from 30+ facilities across 8 U.S. climate zones are reported to achieve 94% of centralized performance, under 1 MB total communication, and <0.15% excess risk from privacy.

Significance. If the simulation fidelity claims hold, the work offers a practical route to cross-facility knowledge transfer in CEA without exposing proprietary grow recipes, potentially delivering the cited 30-38% HVAC energy reductions and faster commissioning. The explicit alignment of the three-tier decomposition with the underlying thermodynamics and cultivar physics, together with the observation that the small PID parameter count renders privacy essentially free, are genuine strengths that distinguish the approach from generic federated baselines.

major comments (2)

[Simulation experiments (as described in the abstract and experimental results)] The central performance claims (94% of centralized training, <1 MB communication, <0.15% excess risk) are obtained exclusively from simulations whose calibration procedure, heterogeneity model, and predictive fidelity are not quantitatively validated. No held-out real-trajectory prediction error, cross-facility transfer error, or comparison of simulated versus logged HVAC/VPD dynamics is reported, which directly affects the reliability of the privacy-utility and heterogeneity-handling claims.
[Framework description and privacy analysis] The statement that privacy is 'essentially for free' (excess risk <0.15%) depends on the specific allocation of tier-specific differential privacy budgets and the 36-parameter PID model size, yet the manuscript provides neither the concrete budget values per tier nor an ablation showing how excess risk scales with those budgets.

minor comments (2)

[Abstract and experimental results] The abstract and results sections would benefit from explicit reporting of error bars, number of random seeds, and at least one ablation (e.g., flat vs. hierarchical federation) to allow readers to assess variability of the 94% figure.
[Model decomposition section] Notation for the three tiers and the PID coefficient vector should be introduced with a single consistent diagram or table early in the paper to improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight important aspects of validation and analysis that we will address to strengthen the manuscript. We respond to each major comment below.

read point-by-point responses

Referee: [Simulation experiments (as described in the abstract and experimental results)] The central performance claims (94% of centralized training, <1 MB communication, <0.15% excess risk) are obtained exclusively from simulations whose calibration procedure, heterogeneity model, and predictive fidelity are not quantitatively validated. No held-out real-trajectory prediction error, cross-facility transfer error, or comparison of simulated versus logged HVAC/VPD dynamics is reported, which directly affects the reliability of the privacy-utility and heterogeneity-handling claims.

Authors: We agree that the manuscript would benefit from explicit quantitative validation of the simulation fidelity to better support the reported performance claims. While the simulations are calibrated on 7+ years of real production data from 30+ facilities, the current text does not include held-out prediction errors or direct comparisons between simulated and logged dynamics. In the revised manuscript we will add a dedicated validation subsection reporting mean-squared and mean-absolute errors on held-out real trajectories for VPD, temperature, and HVAC power, along with cross-facility transfer metrics computed on real data subsets. These additions will directly address the reliability concerns raised. revision: yes
Referee: [Framework description and privacy analysis] The statement that privacy is 'essentially for free' (excess risk <0.15%) depends on the specific allocation of tier-specific differential privacy budgets and the 36-parameter PID model size, yet the manuscript provides neither the concrete budget values per tier nor an ablation showing how excess risk scales with those budgets.

Authors: The referee is correct that the privacy analysis is incomplete without the concrete per-tier budget values and an ablation study. The 36-parameter model size is stated, but the specific differential privacy budgets (ε values) allocated to the global, crop-cluster, and local tiers are not provided, nor is there a scaling analysis. In the revision we will insert a new privacy-analysis subsection that lists the exact tier-specific budgets used to obtain the <0.15% excess-risk figure and includes an ablation plot showing excess risk versus budget allocation. This will make the 'essentially for free' claim fully reproducible and transparent. revision: yes

Circularity Check

0 steps flagged

No circularity: framework and simulation results are independent of self-referential fitting

full rationale

The paper defines HierFedCEA via an explicit three-tier decomposition (global physics, crop-cluster VPD mappings, local equipment adaptation) with tier-specific DP budgets, then reports simulation outcomes on data-calibrated models. No equations reduce the 94% centralized performance, sub-1 MB communication, or <0.15% excess risk figures to quantities defined by the same fitted parameters. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing premises. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 1 invented entities

The central claim rests on the assumption that a 36-parameter PID model can be decomposed into universal physics, crop-cluster, and facility-specific tiers while differential privacy adds negligible excess risk.

free parameters (2)

tier-specific differential privacy budgets
Chosen per tier to achieve privacy essentially for free; values not specified but directly affect the <0.15% excess risk claim.
36-parameter PID model coefficients
Compact model size is leveraged for low communication; coefficients are learned but count as fitted within the federated process.

axioms (2)

domain assumption Universal thermodynamic relationships can be captured in a single global tier independent of crop or facility
Invoked to justify the three-tier decomposition of the neural network PID auto-tuning model.
domain assumption Crop cultivars can be clustered by VPD-to-gain mappings that generalize across facilities
Basis for the crop-cluster tier.

invented entities (1)

three-tier hierarchical decomposition (global physics, crop-cluster, local personalization) no independent evidence
purpose: To align model structure with physical control problem and enable tier-specific privacy
New structure introduced by the framework; no independent evidence outside the paper's simulations.

pith-pipeline@v0.9.0 · 5529 in / 1395 out tokens · 52063 ms · 2026-05-10T13:21:19.427785+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

35 extracted references · 6 canonical work pages · 2 internal anchors

[1]

Controlled environment agriculture market size, share & trends analysis report,

Research Nester, “Controlled environment agriculture market size, share & trends analysis report,” 2025

2025
[2]

IOGRUCloud: A Scalable AI-Driven IoT Platform for Climate Control in Controlled Environment Agriculture

A. Vakhnovskyi, “IOGRUCloud: A scalable AI-driven IoT platform for climate control in controlled environment agriculture,”arXiv preprint arXiv:2604.07586, 2026

work page internal anchor Pith review Pith/arXiv arXiv 2026
[3]

Plant factories versus greenhouses: Comparison of resource use efficiency,

L. Graamanset al., “Plant factories versus greenhouses: Comparison of resource use efficiency,”Agricultural Systems, vol. 160, pp. 31–43, 2018

2018
[4]

Farmer and their data: An examination of farmers’ reluctance to share their data through the lens of the laws impacting smart farming,

L. Wiseman, J. Sanderson, A. Zhang, and E. Jakku, “Farmer and their data: An examination of farmers’ reluctance to share their data through the lens of the laws impacting smart farming,”NJAS-Wageningen J. Life Sciences, vol. 90–91, p. 100301, 2019

2019
[5]

Deep leakage from gradients,

L. Zhu, Z. Liu, and S. Han, “Deep leakage from gradients,” inProc. NeurIPS, 2019

2019
[6]

Communication-efficient learning of deep networks from decentralized data,

B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inProc. AISTATS, 2017

2017
[7]

Federated learning with non-iid data,

Y . Zhaoet al., “Federated learning with non-IID data,”arXiv preprint arXiv:1806.00582, 2018

work page arXiv 2018
[8]

Local model poisoning attacks to Byzantine-robust federated learning,

M. Fang, X. Cao, J. Jia, and N. Gong, “Local model poisoning attacks to Byzantine-robust federated learning,” inProc. USENIX Security, 2020

2020
[9]

Federated deep reinforcement learning for multi-zone HV AC control,

M. Xiaet al., “Federated deep reinforcement learning for multi-zone HV AC control,”IEEE Trans. Smart Grid, 2025

2025
[10]

Federated learning for autonomous HV AC schedul- ing in commercial buildings,

M. Hagstromet al., “Federated learning for autonomous HV AC schedul- ing in commercial buildings,”Energy and Buildings, 2025

2025
[11]

One for many: Transfer learning for building HV AC control,

Z. Xuet al., “One for many: Transfer learning for building HV AC control,” inProc. ACM BuildSys, 2020. DOI: 10.1145/3408308.3427617

work page doi:10.1145/3408308.3427617 2020
[12]

Internet of things for the future of smart agriculture: A comprehensive survey of emerging technologies,

O. Frihaet al., “Internet of things for the future of smart agriculture: A comprehensive survey of emerging technologies,”IEEE/CAA J. Au- tomatica Sinica, vol. 8, no. 4, 2021

2021
[13]

Federated learning for agricultural privacy,

R. Daraet al., “Federated learning for agricultural privacy,”Computers and Electronics in Agriculture, 2025

2025
[14]

Federated learning for Internet of Things: A comprehensive survey,

D. C. Nguyenet al., “Federated learning for Internet of Things: A comprehensive survey,”IEEE Internet of Things Journal, vol. 8, no. 21, pp. 16243–16266, 2021

2021
[15]

Federated learning with non-IID data: A survey,

W. Luet al., “Federated learning with non-IID data: A survey,”IEEE Internet of Things Journal, vol. 11, no. 12, 2024

2024
[16]

A survey on federated learning for resource-constrained IoT devices,

A. Imteajet al., “A survey on federated learning for resource-constrained IoT devices,”IEEE Internet of Things Journal, vol. 9, no. 1, 2022

2022
[17]

Communication-efficient federated learning for wireless edge intelligence in IoT,

J. Mills, J. Hu, and G. Min, “Communication-efficient federated learning for wireless edge intelligence in IoT,”IEEE Internet of Things Journal, vol. 7, no. 7, 2020

2020
[18]

A systematic re- view of real-world deployed machine learning-based HV AC con- trollers,

M. K. Mulayim, F. Schwenker, and M. Beigl, “A systematic re- view of real-world deployed machine learning-based HV AC con- trollers,”Energy and Buildings, vol. 312, p. 114189, 2024. DOI: 10.1016/j.enbuild.2024.114189

work page doi:10.1016/j.enbuild.2024.114189 2024
[19]

Deep reinforcement learning with robust op- timization for greenhouse climate control,

A. Ajagekaret al., “Deep reinforcement learning with robust op- timization for greenhouse climate control,”Computers & Chemical Engineering, 2023

2023
[20]

Adaptive knowledge transfer for federated greenhouse vegetable disease recognition,

Z. Chenet al., “Adaptive knowledge transfer for federated greenhouse vegetable disease recognition,”Frontiers in Plant Science, 2025

2025
[21]

Federated learning for weed detection in precision agriculture,

K. Zhanget al., “Federated learning for weed detection in precision agriculture,”Computers and Electronics in Agriculture, 2024

2024
[22]

Plant responses to rising vapor pressure deficit,

C. Grossiordet al., “Plant responses to rising vapor pressure deficit,” New Phytologist, vol. 226, no. 6, pp. 1550–1566, 2020

2020
[23]

Advances in greenhouse automation and controlled environment agriculture: A transition to plant factories and urban farming,

R. R. Shamshiriet al., “Advances in greenhouse automation and controlled environment agriculture: A transition to plant factories and urban farming,”Int. J. Agricultural and Biological Engineering, vol. 11, no. 1, pp. 1–22, 2018. DOI: 10.25165/j.ijabe.20181101.3210

work page doi:10.25165/j.ijabe.20181101.3210 2018
[24]

Federated optimization in heterogeneous networks,

T. Liet al., “Federated optimization in heterogeneous networks,” inProc. MLSys, 2020

2020
[25]

SCAFFOLD: Stochastic controlled averaging for fed- erated learning,

S. P. Karimireddy, S. Kale, M. Mohri, S. J. Reddi, S. Stich, and A. T. Suresh, “SCAFFOLD: Stochastic controlled averaging for fed- erated learning,” inProc. ICML, 2020

2020
[26]

Federated Learning with Personalization Layers

A. Arivazhaganet al., “Federated learning with personalization layers,” arXiv preprint arXiv:1912.00818, 2019

work page internal anchor Pith review arXiv 1912
[27]

FedBN: Federated learning on non-IID features via local batch normalization,

X. Li, M. Jiang, X. Zhang, M. Kamp, and Q. Dou, “FedBN: Federated learning on non-IID features via local batch normalization,” inProc. ICLR, 2021

2021
[28]

Personalized federated learning with Moreau envelopes,

A. Fallah, A. Mokhtari, and A. Ozdaglar, “Personalized federated learning with Moreau envelopes,” inProc. NeurIPS, 2020

2020
[29]

FLTrust: Byzantine-robust federated learning via trust bootstrapping,

X. Caoet al., “FLTrust: Byzantine-robust federated learning via trust bootstrapping,” inProc. NDSS, 2021

2021
[30]

An efficient framework for clustered federated learning,

A. Ghosh, J. Chung, D. Yin, and K. Ramchandran, “An efficient framework for clustered federated learning,” inProc. NeurIPS, 2020

2020
[31]

Learning differentially private recurrent lan- guage models,

H. B. McMahanet al., “Learning differentially private recurrent lan- guage models,” inProc. ICLR, 2018

2018
[32]

Private empirical risk mini- mization: Efficient algorithms and tight error bounds,

R. Bassily, A. Smith, and A. Thakurta, “Private empirical risk mini- mization: Efficient algorithms and tight error bounds,” inProc. FOCS, 2014

2014
[33]

K. J. ˚Astr¨om and T. H ¨agglund,Advanced PID Control. ISA, 2006. IEEE INTERNET OF THINGS JOURNAL 7

2006
[34]

dp-accounting: Differential privacy accounting library,

Google, “dp-accounting: Differential privacy accounting library,” 2023. [Online]. Available: https://github.com/google/differential-privacy

2023
[35]

Distributed HV AC control with edge computing,

Y . Suet al., “Distributed HV AC control with edge computing,”IEEE Internet of Things Journal, vol. 9, 2022. Andrii Vakhnovskyireceived the B.S. degree in computer engineering and the M.S. degree in systems engineering from the National Technical University “Kharkiv Polytechnic Institute” (NTU “KhPI”), Ukraine, in 2009 and 2011, respectively. He is the Fo...

2022