⚡ electrical engineering

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

HierFedCEA is a novel hierarchical federated learning framework that enables privacy-preserving, energy-efficient climate control optimization across diverse Controlled Environment Agriculture facilities by decomposing a compact PID model into three specialized tiers, thereby allowing operators to share knowledge without exposing sensitive grow recipes while achieving near-centralized performance with minimal communication overhead.

Original authors: Andrii Vakhnovskyi

Published 2026-04-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Andrii Vakhnovskyi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the manager of a high-tech greenhouse. You have spent years and millions of dollars figuring out the perfect recipe to grow the juiciest tomatoes or the most potent cannabis. You know exactly how much heat, humidity, and light your plants need at every stage of their life.

Now, imagine you want to open a new greenhouse in a different city. You could start from scratch, spending months guessing what works, or you could ask your friends who run other greenhouses for their advice.

The Problem: Your friends (other facility owners) are terrified to share their secrets. They think, "If I tell you my exact temperature settings, you'll steal my recipe and sell it cheaper than me!" Also, if you put all your data into one big cloud computer, hackers might steal it, or a competitor might see it.

The Solution: This paper introduces HierFedCEA, a clever new way for greenhouses to learn from each other without ever sharing their actual data.

Here is how it works, using a simple analogy:

The "Secret Recipe" Analogy

Think of controlling a greenhouse like baking a cake. To bake the perfect cake, you need three things:

Universal Physics: The laws of baking (e.g., "heat makes things rise," "sugar caramelizes"). This is the same whether you are in New York or California.
The Specific Flavor: The recipe for a chocolate cake vs. a vanilla cake. This depends on the crop (tomatoes vs. lettuce).
The Specific Oven: Your specific oven might run 5 degrees hotter than your neighbor's, or it might take longer to preheat. This is unique to your building.

Old Way (Centralized Learning): Everyone sends their entire recipe book to a central chef. The chef mixes them all together and sends a new book back.

Risk: Everyone sees everyone else's secret recipes.

The HierFedCEA Way (Hierarchical Federated Learning):
Instead of sending the whole book, the system splits the learning into three separate "layers" and only shares what is safe to share.

Layer 1: The Universal Laws (Shared with Everyone)

The system shares the physics. "Heat rises," "humidity condenses."

Analogy: Everyone agrees that "fire makes things hot." This is true for a cannabis farm in Arizona and a lettuce farm in Illinois.
Privacy: This is public knowledge, so sharing it is safe.

Layer 2: The Crop Family (Shared with Similar Farmers)

The system groups farmers by what they grow. All the "Lettuce Farmers" talk to each other. All the "Tomato Farmers" talk to each other.

Analogy: The lettuce farmers share tips on how to keep lettuce crisp, but they don't tell the tomato farmers how to grow tomatoes.
Privacy: You only share your "crop secrets" with your specific crop family, not the whole world.

Layer 3: The Personal Touch (Kept Private)

The system keeps the specific equipment settings completely private.

Analogy: "My oven runs hot, so I turn the dial down by 2 degrees." This is your unique quirk. You never tell anyone this. You just adjust your own oven based on what you learned from the group.

Why is this a Big Deal?

It's Super Fast: Because the system separates the "universal stuff" from the "specific stuff," it learns 3.6 times faster than other methods. A new greenhouse can go from "cold start" to "perfectly optimized" in just two weeks instead of six months.
It's Private: The paper proves that even if you try to hack the system to steal the data, the math makes it impossible to reconstruct the secret recipes. It's like trying to guess a specific cake recipe just by looking at the temperature of the kitchen.
It's Tiny: The data being shared is so small (less than 1 Megabyte total for the whole group) that it doesn't clog up the internet, even for slow rural connections.

The Result

By using this "Hierarchical" (layered) approach, the greenhouses get 94% of the benefits of having one giant super-computer with all the data, but 100% of the privacy.

In short: HierFedCEA is like a secret society of farmers who can whisper their best tips to each other in groups, without ever revealing their most guarded secrets, ensuring everyone grows better crops while saving massive amounts of energy.

1. Problem Statement

Controlled Environment Agriculture (CEA) facilities (greenhouses, vertical farms) rely on complex HVAC systems to regulate temperature, humidity, and CO2. Optimizing these systems using AI (specifically Neural Network-based PID auto-tuning) can reduce energy consumption by 30–38%. However, current optimization methods face a critical bottleneck: data privacy.

The Conflict: Facility operators possess valuable, proprietary "grow recipes" (R&D investments worth millions) encoded in their operational data. They are unwilling to share raw sensor data with a central cloud server due to competitive concerns and the risk of gradient inversion attacks.
The Challenge: Standard Federated Learning (FL) approaches fail in CEA due to:
- Extreme Data Heterogeneity: Different crops (e.g., cannabis vs. lettuce) have vastly different optimal control strategies (non-IID data).
- Equipment Heterogeneity: HVAC systems from 50+ manufacturers have different actuator dynamics.
- Small Federation Size: With only 5–30 participants, individual clients have outsized influence, increasing vulnerability to attacks and convergence issues.
- Temporal Non-stationarity: Crop growth stages and seasonal changes cause concept drift.

2. Methodology: HierFedCEA

The authors propose HierFedCEA, a hierarchical federated learning framework that decomposes the neural network PID auto-tuning model based on physical meaning rather than just network topology.

A. Model Architecture (3-Tier Decomposition)

The core innovation is splitting the 36-parameter Multi-Layer Perceptron (MLP) into three distinct tiers:

Tier 1: Global Physics (18 parameters)
- Scope: Captures universal thermodynamic relationships (heat transfer, psychrometrics) governed by laws like Clausius-Clapeyron.
- Sharing: Aggregated across all facilities globally.
- Privacy: Low sensitivity; uses a higher privacy budget ( $\epsilon = 8$ ).
Tier 2: Crop Cluster (17 parameters)
- Scope: Encodes crop-specific Vapor Pressure Deficit (VPD) sensitivity and gain-scheduling.
- Sharing: Aggregated only within crop-family clusters (e.g., all cannabis facilities, all leafy greens).
- Privacy: High sensitivity; uses a stricter privacy budget ( $\epsilon = 4$ ).
- Clustering: Facilities are initially grouped by crop type and refined dynamically using gradient cosine similarity.
Tier 3: Local Personalization (1 parameter)
- Scope: Adapts to facility-specific equipment dynamics (actuator dead time and lag).
- Sharing: Never shared. This parameter remains local and is updated online.
- Privacy: Maximum protection ( $\epsilon = \infty$ ).

B. Aggregation Protocol

Intra-Cluster Aggregation: Tier 2 updates are aggregated within crop clusters using data-proportional weighted averaging with local differential privacy (Gaussian noise).
Global Aggregation: Tier 1 updates are aggregated globally using a trust-weighted mechanism (inspired by FLTrust) to detect Byzantine attacks, combined with FedProx regularization to handle heterogeneity.
Cold-Start Protocol: New facilities join by downloading the Global and Cluster models (35/36 parameters) and learning only the local parameter, enabling rapid convergence.

C. Privacy Analysis

The authors demonstrate that the extreme compactness of the 36-parameter model makes Differential Privacy (DP) "essentially free."

Theoretical analysis shows that at $\epsilon = 4$ , the excess empirical risk is < 0.15%.
This contrasts sharply with large vision/language models where DP noise significantly degrades performance.

3. Key Contributions

Physics-Informed Decomposition: The first FL framework for CEA that splits model parameters based on physical domain knowledge (thermodynamics vs. crop biology vs. equipment dynamics) rather than neural network layers.
Privacy-Utility Efficiency: Proves that for compact control models, strong privacy guarantees ( $\epsilon=4$ ) incur negligible utility loss (<0.15% risk), solving the privacy-utility tradeoff.
Tier-Specific Privacy & Cadence: Implements different privacy budgets and communication frequencies tailored to the sensitivity and timescale of each physical phenomenon.
Comprehensive Evaluation: Validated via simulation calibrated with 7+ years of production data from 30+ commercial facilities across 8 U.S. climate zones.

4. Experimental Results

The system was evaluated against 8 baselines (including Local-Only, Centralized, FedAvg, FedProx, SCAFFOLD, etc.) over a 180-day simulation.

Performance: HierFedCEA achieved 94% of the performance of a centralized training system (which violates privacy) and significantly outperformed standard FL methods.
- VPD Tracking RMSE: 0.068 kPa (vs. 0.098 kPa for FedAvg and 0.142 kPa for Local-Only).
- Energy Savings: 34.1% reduction in HVAC energy consumption.
Convergence Speed:
- Rounds to Convergence: 52 rounds (vs. 187 for FedAvg).
- Improvement: 3.6× faster convergence than flat FedAvg due to reduced gradient dissimilarity within tiers.
Robustness to Heterogeneity: Under high heterogeneity (mixed crops/climates), HierFedCEA's performance degraded by only 12%, whereas FedAvg degraded by 47%.
Cold-Start: New facilities reached 85% of fleet-average performance in 14 days (vs. 28 days for FedAvg and 56+ days for Local-Only).
Communication Cost: Total communication for the entire federation over 100 rounds was < 1 MB (approx. 840 KB), making it viable for low-bandwidth IoT connections.

5. Significance

First of its Kind: This is the first federated learning framework specifically designed for CEA climate control, addressing a gap where FL has been applied to building HVAC but not the unique biological constraints of agriculture.
Commercial Viability: By enabling cross-facility knowledge transfer without exposing proprietary grow recipes, it unlocks the potential for industry-wide AI optimization, potentially saving the sector millions in energy costs and R&D time.
Scalability: The framework demonstrates that FL is feasible for small, highly heterogeneous industrial IoT fleets, provided the model architecture aligns with the underlying physics of the problem.
Privacy Guarantee: It offers a practical solution to the "data silo" problem in agriculture, allowing operators to collaborate without compromising trade secrets.