Smart Thermostat IoT

Smart Thermostat IoT — Topics for IoT Students

IEEE ACCESS 1

IOGRUCloud: A Scalable AI-Driven IoT Platform

for Climate Control in Controlled Environment Agriculture ANDRII VAKHNOVSKYI IOGRU LLC, New York, NY 10022, USA

Abstract—Climate control in Controlled Environment Agri- for climate regulation accounts for 20–50% of CEA operating arXiv:2604.07586v1 [eess.SY] 8 Apr 2026

culture (CEA) remains dominated by static-setpoint controllers budgets, with HVAC systems alone consuming 30–80% of with isolated PID loops, leading to cross-coupling conflicts, total facility energy depending on geographic location and suboptimal energy consumption, and limited scalability across heterogeneous facilities. This paper presents IOGRUCloud, a outdoor climate conditions [3], [4]. three-tier IoT platform for AI-driven climate control deployed Conventional CEA climate control relies on static-setpoint across 30+ commercial CEA facilities in 8 U.S. climate zones controllers with isolated PID loops for individual parame- over 7+ years of continuous operation (2017–2024) — the ters [5].

This approach suffers from fundamental limitations: largest documented real-world deployment of adaptive climate (1) static setpoints do not adapt to changing outdoor condi- control in this domain, exceeding the combined duration of all published field experiments in AI-based HVAC control by tions, leading to suboptimal energy consumption; (2) indepen- approximately 60 times. The platform introduces a cascading dent PID loops create cross-coupling conflicts — for example, control architecture where Vapor Pressure Deficit serves as the simultaneous heating and dehumidification operations that primary setpoint, with a neural network optimizer selecting work against each other; (3) reactive control responds to energy-minimal temperature–humidity combinations on the VPD deviations after they occur rather than anticipating them; and constraint surface.

Inner PID loops with neural network self- tuning track the resulting setpoints at higher bandwidth. A four- (4) the lack of cross-subsystem coordination results in energy level progressive autonomy model (L1–L4) enables graduated waste from contradictory actuator commands [6]. transition from anomaly detection to autonomous optimization Recent advances in artificial intelligence have shown with operator guardrails. The edge-first architecture ensures all promise for CEA climate control.

Chen et al. [7] review AI control logic executes locally without cloud dependency, while applications in CEA, identifying deep reinforcement learn- the cloud tier enables cross-facility transfer learning across the fleet. Multi-facility aggregate results demonstrate 30–38% HVAC ing and model predictive control as leading approaches. energy reduction, 68–73% improvement in VPD stability, and Ajagekar et al. [8] demonstrated up to 57% energy reduction 60–67% faster disturbance recovery compared to conventional using deep reinforcement learning with robust optimization controllers.

Two detailed case studies — a 40,000 sq ft desert- — however, this result was obtained in simulation on a climate facility and a 120,000 sq ft continental-climate facility single greenhouse model. Adesanya et al. [9] applied deep — are presented with quantitative metrics. Practical deployment lessons from integrating 50+ equipment manufacturers via 8 in- reinforcement learning for PID parameter tuning in greenhouse dustrial protocols and achieving 1–5 day commissioning timelines HVAC, while Panagopoulos et al. [10] proposed a cascaded are discussed. economic model predictive control approach for greenhouse Index Terms—Cascading control, controlled environment agri- climate — both validated only in simulation.

The iGrow sys- culture, edge computing, energy optimization, HVAC, Internet tem [11] demonstrated reinforcement learning for autonomous of Things (IoT), neural network, PID controller, progressive greenhouse control with measurable yield improvements, but autonomy, scalable deployment, vapor pressure deficit. at limited scale. The hybrid approach combining classical PID control I. I NTRODUCTION with neural network adaptation has been explored since the work of Zeng et al. [12], who proposed RBF neural

T HE global Controlled Environment Agriculture (CEA)

industry is undergoing rapid expansion as climate change, urbanization, and food security concerns drive demand for network-augmented PID for greenhouse temperature regula- tion. Salehi et al. [13] validated neural network-based PID auto-tuning in an industrial setting, demonstrating production- year-round, location-independent crop production [1], [2]. scale reliability. The concept of progressive industrial auton- CEA facilities — including greenhouses, vertical farms, and omy has been formalized by Gamer et al. [14], who proposed a indoor cultivation operations — require precise simultane- six-level taxonomy for autonomous industrial plants.

The IoT ous management of dozens of interdependent environmental and edge-cloud computing paradigm for agricultural applica- parameters: air temperature, relative humidity, CO2 concen- tions has been established through reference architectures by tration, photosynthetically active radiation (PAR), substrate Alreshidi [15] and Sami and Ibraheem [16]. The physiological moisture, and nutrient solution chemistry. Energy expenditure basis for using Vapor Pressure Deficit (VPD) as a primary Corresponding author: Andrii Vakhnovskyi (e-mail: control variable is well-established: Grossiord et al. [17] andrii.vakhnovskyi@gmail.com).

ORCID: 0009-0007-8306-5932. demonstrated that VPD is a critical driver of plant stress

IEEE ACCESS 2

responses, while Inoue et al. [18] showed that minimizing VPD fluctuations directly improves plant growth. Despite these advances, a critical gap persists between simulation-validated research and production-scale deploy- ment. Mulayim et al. [19] found that the combined dura- tion of all peer-reviewed real-world field experiments in AI- based HVAC control totals approximately 43 days globally.

Al Sayed et al. [20] reported that only 23% of reinforcement learning studies for HVAC involved real buildings. The largest published real-world deployment — Moshari et al. [21] with 13 buildings over one heating season — remains limited in both facility count and temporal scope. Furthermore, no commercial CEA control platform — in- cluding Priva, Argus Controls, TrolMaster, or Wadsworth — has published its system architecture, control algorithms, or deployment-scale performance data in peer-reviewed litera- Fig. 1.

Three-tier architecture diagram: Field Layer → Edge AI Layer → ture [5]. The only comparable end-to-end agricultural IoT Cloud Layer. platform paper is FarmBeats [22], which describes a monitor- ing and analytics platform (not a closed-loop control system) deployed at 2 farms. Fig. 1.

This paper addresses these gaps. The main contributions are as follows: A. Field Layer (Distributed Sensor Network) 1) We present the architecture of IOGRUCloud, a three-tier IoT platform for CEA climate control deployed across The field layer comprises a distributed network of industrial 30+ commercial facilities in 8 U.S. climate zones — the sensors and actuators communicating via standard protocols. largest peer-reviewed deployment of its kind, exceeding Table I lists the sensor types and specifications deployed in a all published agricultural IoT field experiments by an typical CEA facility zone. order of magnitude.

A typical multi-zone facility (10–30 zones) generates 300– 2) We propose a cascading VPD control architecture where 1,800 data points at 1-second polling frequency, with storage Vapor Pressure Deficit serves as the primary setpoint, at 10-second resolution. Large multi-state operator facilities and a neural network optimizer determines the energy- generate up to 3,000+ data points. The platform integrates minimal temperature and humidity targets on the VPD equipment from 50+ manufacturers — including Carrier, constraint surface.

Trane, Daikin, LG, and Mitsubishi for HVAC; Quest, Anden, 3) We describe a neural network-based PID self-tuning and Desert Aire for dehumidification; and Fluence, Gavita, mechanism with Lyapunov stability guarantees, de- and Growers Choice for lighting — through a vendor-agnostic ployed in production HVAC systems — the first reported integration layer spanning 8 industrial protocols: BACnet/IP, real-world deployment of neural network-tuned PID Modbus RTU/TCP, MQTT, OPC UA, 0–10 V analog, 4– controllers in CEA. 20 mA analog, SDI-12, and REST/WebSocket APIs. 4) We introduce a four-level progressive autonomy model Multi-sensor redundancy within each zone employs median (L1–L4) for CEA automation, with integrated confi- filtering across co-located sensors of the same type to de- dence scoring and operator guardrails. tect and reject sensor drift without corrupting the aggregate 5) We report production deployment results including 30– measurement [23].

When the z-score of any individual sensor 38% HVAC energy reduction and practical lessons from reading exceeds 2.5 relative to its co-located peers, the reading integrating 50+ equipment manufacturers via 8 industrial is flagged for maintenance review and excluded from the protocols with 1–5 day commissioning timelines. control signal. The remainder of this paper is organized as follows. Sec- tion II describes the three-tier system architecture.

Section III B. Edge AI Layer (Local Decision-Making) presents the cascading VPD control approach and neural network PID self-tuning with stability analysis. Section IV The edge controller is the central architectural element.

All introduces the progressive autonomy model. Section V de- real-time control logic executes locally on industrial ARM/x86 scribes the multi-facility deployment methodology. Section VI hardware without cloud dependency — a critical requirement presents experimental results from 30+ facilities with two de- for commercial agricultural facilities where network outages tailed case studies.

Section VII discusses practical deployment must not disrupt climate control. lessons. Section VIII concludes the paper. The edge layer hosts: 1) The cascading VPD controller with PID auto-tuning II.

S YSTEM A RCHITECTURE (Section III). The IOGRUCloud platform implements a three-tier archi- 2) The neural network AI module implementing progres- tecture with clearly separated responsibilities, illustrated in sive autonomy levels L1–L4 (Section IV).

IEEE ACCESS 3

TABLE I

S ENSOR S PECIFICATIONS P ER C ONTROL Z ONE

Parameter Sensor type Accuracy Interface Qty/zone

Air temperature Aspirated climate station ±0.1 ◦ C Modbus RTU 2–4 Relative humidity Aspirated climate station ±1.5% RH Modbus RTU 2–4 CO2 concentration NDIR sensor ±50 ppm Modbus / 4–20 mA 1–2 PAR/PPFD Quantum sensor (400–700 nm) ±5% SDI-12 / 4–20 mA 1–2 Substrate VWC Capacitance probe ±3% SDI-12 3–8 Substrate EC Capacitance probe ±10% SDI-12 3–8 Substrate temperature Thermistor ±0.5 ◦ C SDI-12 3–8 Solution EC (in/out) Inline transmitter ±2% 4–20 mA / Modbus 2 Solution pH (in/out) ISFET / glass electrode ±0.02 4–20 mA / Modbus 2 Water flow Pulse flow meter ±2% Pulse 1–4 Differential pressure Transmitter ±1 Pa 4–20 mA 1 Power consumption CT clamp per circuit ±1% 4–20 mA 4–16 Leaf temperature IR thermometer ±0.5 ◦ C Modbus 1–2

3) Recipe management and stage-based automation with configurable IF/THEN/ELSE logic. 4) Safety systems including hardware watchdog timers, failsafe modes, and equipment interlock logic. Local data storage uses TimescaleDB for time-series data at 10-second resolution (supporting 2–8 years of retention) and SQLite for configuration. The air-gapped failsafe mode ensures continued operation during network outages: the edge controller maintains full autonomous control capability using locally cached recipes, setpoints, and learned parameters.

Typical edge configuration: 1 controller per 5–10 zones.

C. Cloud Layer (Fleet Learning)

The cloud tier aggregates anonymized operational data from the facility network and enables capabilities that require cross- Fig. 2. Cascading VPD control block diagram. facility visibility: 1) Cross-facility model training — learned operational patterns from established facilities are used to accel- commands — the system targets VPD directly and uses a erate optimization at new deployments. Optimal VPD neural network optimizer to decompose VPD targets into trajectories for specific crop stages, equipment-specific energy-optimal temperature–humidity combinations.

PID tuning parameters, and seasonal control strategy The cascading architecture operates as follows (Fig. 2): templates are transferred through the fleet [24]. The outer loop (AI optimizer) operates on VPD error and 2) Multi-facility benchmarking — facilities operating sim- selects the energy-minimal point on the VPD constraint surface ilar crops under comparable recipes are compared to in T –RH space. The inner loops (PID controllers) track identify optimization opportunities. the resulting temperature and humidity setpoints.

The inner 3) Digital twin simulation — proposed recipe modifications loops operate at 3–10× higher bandwidth than the outer loop, are simulated against historical data before application satisfying the cascade speed separation requirement [10]. to live facilities [25]. The saturation vapor pressure is computed using the Tetens 4) Real-time monitoring and alarming — sub-second equation with improved coefficients from Alduchov and Es- telemetry across all deployed sites with centralized op- kridge [26]: erational analytics.   The cloud layer supports L4-level autonomous optimization 17.625 × T es (T ) = 0.61094 × exp (1) (Section IV), including predictive yield modeling and energy T + 243.04 cost optimization via load shifting to off-peak tariff hours. where es is in kPa and T in ◦ C.

This improved Magnus form yields relative errors below 0.01% in the 0–50 ◦ C range. III. AI-D RIVEN C LIMATE C ONTROL The Vapor Pressure Deficit is: A.

Cascading VPD Control Architecture   RH The central innovation in control strategy is the elevation of VPD = es (Tair ) × 1 − (2) VPD from a monitored metric to the primary cascading control 100 variable. Rather than independently regulating temperature For facilities equipped with infrared canopy temperature and humidity — which often produces contradictory actuator sensors:

IEEE ACCESS 4

Initial gain estimation uses the Ziegler–Nichols relay feed-

RH VPDleaf = es (Tleaf ) − es (Tair ) × (3) back method [28]: a relay of amplitude ±h is applied to the 100 feedback loop, producing sustained oscillations at the ultimate The system supports three VPD computation modes: air frequency. The ultimate gain Ku = 4h/(πa) and ultimate VPD (standard), leaf VPD (with canopy sensors), and canopy period Tu are extracted, yielding initial gains Kp = 0.6 Ku , VPD (multi-point spatial average) [17], [27].

Ti = Tu /2, Td = Tu /8. Typical auto-tuning time: 15–30 minutes per control loop. B.

Energy-Optimal VPD Decomposition Online adaptation is performed by a three-layer backprop- agation neural network: The VPD constraint es (T ∗ ) × (1 − RH∗ /100) = VPDtarget defines a curve in T –RH space. Any point on this curve satisfies the VPD requirement. The system selects the energy-  Input: x(k) = e(k), e(k−1), e(k−2), ∆e(k), minimal point by solving:  SP(k), PV(k), u(k−1) (9)

min E(T ∗ , RH∗ ) = αh · max(T ∗ − Tcur , 0) The hidden layer consists of 3 neurons with sigmoid acti- T ∗ ,RH∗ vation. The output layer produces 3 neurons: + αc · max(Tcur − T ∗ , 0) + αd · max(RHcur − RH∗ , 0)    Kp (k), Ki (k), Kd (k) = σ(r1 ) · Kp,max , + αm · max(RH∗ − RHcur , 0) (4) σ(r2 ) · Ki,max , subject to:  σ(r3 ) · Kd,max (10) ∗ ∗ es (T ) × (1 − RH /100) = VPDtarget where σ is the sigmoid function and Kn,max defines the ∗ Tmin ≤ T ≤ Tmax , ∗ RHmin ≤ RH ≤ RHmax feasible gain range.

The weight update follows backprop- agation with objective E(k) = 21 [SP(k) − PV(k)]2 , using where αh , αc , αd , αm are energy cost coefficients for sign(∂y/∂u) to approximate the unknown plant Jacobian. heating, cooling, dehumidification, and humidification respec- The static+dynamic decomposition ensures baseline stability: tively. Ziegler–Nichols initial gains provide the stable operating point, Using the VPD constraint, the problem is parameterized by while the neural network applies incremental corrections [12], temperature: [13].   VPDtarget RH∗ (T ) = 100 × 1 − (5) es (T ) D.

Lyapunov Stability Analysis reducing the optimization to one dimension: For industrial deployment, stability of the adaptive system must be guaranteed. Consider the Lyapunov function candi- min E(T, RH∗ (T )) (6) date: T ∈[Tmin ,Tmax ]

This 1D optimization is solved at each control cycle by 1

V (k) = e(k)2 (11) the neural network optimizer, which learns the energy cost 2 surface from operational data. The sensitivity analysis yields The change at each step is: ∂VPD/∂T ≈ 2–3 × ∂VPD/∂RH at typical CEA temper- atures (20–30 ◦ C), meaning temperature adjustments have a disproportionately larger effect on VPD — an asymmetry the 1 e(k+1)2 − e(k)2  ∆V (k) = V (k+1) − V (k) = (12) optimizer exploits for energy-efficient control. 2 The adaptive learning rate η is constrained such that C.

Neural Network PID Self-Tuning ∆V (k) < 0 for all k where e(k) ̸= 0. This is achieved The inner-loop PID controllers use the velocity (incremen- by bounding the neural network weight updates so that gains tal) form: Kp (k), Ki (k), Kd (k) remain within the stability region de- fined by the Ziegler–Nichols initial parameters ±∆max . The L3   autonomy guardrails (Section IV) provide a physical-domain ∆u(k) = Kp (k) e(k) − e(k−1) + Ki (k) e(k) manifestation of this constraint: the AI may modify setpoints   + Kd (k) e(k) − 2e(k−1) + e(k−2) (7) only within proven-safe envelopes.

u(k) = u(k−1) + ∆u(k) (8) E. Multi-Objective Optimization where e(k) = SP(k) − PV(k). Anti-windup is implemented The overall optimization integrates VPD tracking accuracy, via conditional integration when the output saturates [28]. energy cost, and equipment preservation:

IEEE ACCESS 5

Photoperiod and stage transitions are prohibited at L3 as

2 they affect plant physiology irreversibly. Every autonomous R(st , at ) = −w1 VPDtarget − VPDmeasured action is logged with timestamp, parameter, old/new values, − w2 Ecost (at ) − w3 Wequip (at ) reason, and confidence score, providing full auditability and X
2 undo capability. − w4 max 0, gi (st ) (13) i

where w1 . . . w4 are weighting coefficients, Ecost represents D. Level L4 — Full Autonomous Optimization electricity cost proportional to actuator power, Wequip penalizes Activated only after ≥3 complete growth cycles with docu- rapid actuator switching to reduce mechanical wear (address- mented outcomes. L4 capabilities include digital twin simula- ing the short-cycling problem with minimum compressor run tion for recipe testing before application [25], predictive yield times of 5 minutes and minimum off times of 3 minutes [1]), modeling, energy cost optimization through load shifting to and the constraint violation terms gi enforce temperature, off-peak tariff hours, and cross-facility knowledge transfer. humidity, and CO2 bounds.

V. M ULTI -FACILITY D EPLOYMENT M ETHODOLOGY IV. P ROGRESSIVE AUTONOMY M ODEL A.

Standardized Onboarding Adapting the industrial autonomy taxonomy of Traditional building management system (BMS) commis- Gamer et al. [14] to the specific requirements of CEA sioning requires weeks to months of on-site engineering. operations, we define four graduated autonomy levels The IOGRUCloud platform achieves 1–5 day commissioning (Table II). through standardization of: 1) BACnet object tables — Pre-built templates for all A.

Level L1 — Anomaly Detection supported HVAC equipment families (Carrier, Trane, Daikin, LG, Mitsubishi) map standard BACnet objects The system monitors all sensor streams and detects de- to the platform’s internal data model. When a facility viations from learned behavior using an autoencoder-based uses supported equipment, integration is configuration anomaly detector trained on 14-day rolling baselines. Anomaly rather than custom engineering. categories include: (1) sensor drift, detected via cross-sensor 2) Unified sequences of operations — Standard control z-score comparison; (2) equipment degradation, identified sequences for common CEA HVAC configurations (split through trending analysis of setpoint achievement time; (3) en- systems, rooftop units, chilled water, VRF) are parame- vironmental anomalies via statistical comparison with rolling terized and instantiated per facility. baselines; and (4) irrigation anomalies through dry-back curve 3) Standardized I/O mapping — A consistent mapping analysis.

At L1, the system generates alerts but takes no framework between physical I/O points and logical autonomous action. control variables eliminates per-facility custom wiring diagrams. B. Level L2 — Recommendations with Confidence Scoring 4) Auto-tuning on first start — The Ziegler–Nichols re- The AI module generates specific corrective recommenda- lay feedback procedure (Section III-C) automatically tions, each with a quantified confidence score.

For example: characterizes each control loop, eliminating manual PID “Reduce night temperature by 1.5 ◦ C — current differential tuning. is insufficient for generative growth at week 5 of flowering. Based on 23 analogous cycles. Confidence: 76%.” Operators B.

Multi-Protocol Integration can accept, dismiss, schedule, or request explanations. Dis- The platform’s vendor-agnostic integration layer abstracts missed recommendations feed back into the learning system, equipment diversity behind a unified internal API. Table IV reducing false positive rates over time. summarizes the protocol coverage.

C. Level L3 — Autonomous Control within Guardrails TABLE IV

P ROTOCOL I NTEGRATION M ATRIX The system autonomously adjusts parameters within operator-defined bounds (Table III). Protocol Equipment category Typical manufacturers BACnet/IP HVAC, BAS Carrier, Trane, Daikin TABLE III Modbus RTU/TCP Sensors, VFDs METER, Gavita, Quest L3 AUTONOMY G UARDRAILS 0–10 V analog LED, VFDs Fluence, Phantom 4–20 mA analog Transmitters EC, pH, pressure, flow Parameter Allowed auto-correction Prohibited actions SDI-12 Substrate METER TEROS 12/21 MQTT IoT devices Custom gateways Temperature ±2 ◦ C from recipe — OPC UA Interoperability Schneider, Siemens Humidity ±5% from recipe — REST/WS Cloud, apps Integrations Irrigation volume ±20% — EC ±0.3 mS/cm from recipe — Photoperiod — All modifications This multi-protocol capability enables the platform to in- Growth stage transitions — All modifications tegrate into existing facilities without requiring equipment

IEEE ACCESS 6

TABLE II

P ROGRESSIVE AUTONOMY L EVELS FOR CEA AUTOMATION

Level Name Capability Action authority

L1 Observation Anomaly detection Alerts only L2 Recommendation Pattern analysis with confidence scoring Operator decides L3 Autonomous (bounded) Setpoint adjustment within guardrails Automated with logging L4 Full optimization Digital twin simulation + cross-facility learning Automated, full scope

replacement — a critical factor for commercial adoption where TABLE V CEA operators have existing capital investments in diverse AVERAGE P ERFORMANCE C OMPARISON ACROSS 30+ FACILITIES (2017–2024) equipment. Metric Before After Change HVAC energy 100% 62–70% −30 to −38% Water use 100% 78–85% −15 to −22% C. Cross-Facility Transfer Learning VPD σ (kPa) 0.15–0.25 0.04–0.08 −68 to −73% Recovery time 12–25 min 4–8 min −60 to −67% Operational patterns learned at established facilities are Alarms/month 8–15 1–3 −75 to −80% Downtime 4.2 h/mo 1.1 h/mo −74% transferred to new deployments through the cloud tier: 1) Optimal VPD trajectories — Per-crop, per-stage VPD The HVAC energy savings of 30–38% are consistent with setpoint curves refined from fleet-wide data. published benchmarks: the best real-world result in the liter- 2) Equipment-specific PID parameters — Learned tuning ature is 35.44% from a single-facility RL deployment [29], parameters for specific equipment models are transferred while simulation-based studies report up to 57% [8].

Our to new installations with the same hardware. results are notable for being achieved across 30+ facilities with 3) Seasonal control templates — Climate-zone-specific heterogeneous equipment, diverse climate zones, and extended seasonal strategies (winter heating optimization, summer temporal scope — conditions that typically degrade control dehumidification) accelerate adaptation. performance compared to controlled single-site experiments. 4) Anomaly detection baselines — Normal operating en- The reduction in critical alarms (75–80%) reflects the sys- velopes established across the fleet improve anomaly tem’s predictive capability: disturbances are anticipated via detection at new sites from day one. weather forecast integration and sensor trend analysis, and preemptively mitigated rather than detected reactively.

The 74% reduction in equipment downtime correlates with the VI. E XPERIMENTAL R ESULTS anomaly detection engine (L1) identifying equipment degra- dation patterns before failure. A.

Deployment Scale and Methodology C. Case Study 1: Desert Climate (Nevada, 40,000 sq ft) The system has been deployed across 30+ commercial CEA facilities in 8 U.S. states spanning diverse climate zones: hot- Table VI presents results from a facility operating in dry desert (Arizona, Nevada), humid subtropical (Florida), Nevada’s extreme hot-dry desert climate. and humid continental (Illinois, Michigan, Ohio, New Jer- sey, Pennsylvania).

Total managed area exceeds 930,000 m2 TABLE VI C ASE S TUDY — N EVADA , 40,000 SQ FT, D ESERT C LIMATE (10,000,000 sq ft) with 500+ individual control zones. Data collection covers 2017–2024 (7+ years of continuous opera- Parameter Before After Change tion). Electricity $18,400/mo $11,700/mo −36.4% The comparative methodology uses a before/after design: Water $3,200/mo $2,450/mo −23.4% each facility’s performance under conventional static-setpoint VPD σ 0.22 kPa 0.06 kPa −72.7% Yield (g/sq ft) 52 61 +17.3% controllers (Priva, Argus, TrolMaster) serves as the baseline, Energy/yield 1.42 kWh/g 0.97 kWh/g −31.7% measured for ≥3 months prior to IOGRUCloud deployment.

Post-deployment metrics are computed over the full opera- tional period. In this environment with extreme diurnal temperature swings (>25 ◦ C day-night differential) and very low ambient humidity (<15% RH), the cascading VPD control provides particular advantage. Rather than maintaining rigid temper- B.

Multi-Facility Aggregate Results ature and humidity setpoints — requiring maximum HVAC capacity during afternoon peaks — the AI optimizer exploits Table V presents aggregate performance metrics averaged the VPD constraint surface. For example, allowing temperature across the deployment fleet. to rise to 25.5 ◦ C while adjusting humidity to 52% achieves

IEEE ACCESS 7

the same target VPD as the conventional 24 ◦ C/55% setpoint, occur frequently in commercial facilities. Facilities that but with significantly reduced cooling load. lost internet connectivity for hours or days continued Predictive control based on weather forecast integration [30] operating normally because all control logic executed further reduces consumption: before anticipated heat waves, locally. The cloud tier proved valuable for optimization the system pre-cools during nighttime off-peak hours, exploit- but not for real-time control. ing the desert’s cool nights and lower electricity tariffs. 3) Standardized BACnet object tables enabled rapid com- missioning.

The pre-built equipment templates reduced D. Case Study 2: Continental Climate (Illinois, 120,000 sq ft) commissioning from weeks to 1–5 days. The most sig- nificant time savings came from eliminating per-facility Table VII presents results from a larger facility in Illinois custom BACnet point mapping. with extreme seasonal variation. 4) Progressive autonomy built operator trust.

Starting at L1 (observation only) and gradually increasing autonomy TABLE VII C ASE S TUDY — I LLINOIS , 120,000 SQ FT, C ONTINENTAL C LIMATE over weeks allowed operators to build confidence in the system. Facilities that attempted to jump directly to L3 Parameter Before After Change autonomous control experienced higher resistance and HVAC (winter) $42,000/mo $27,300/mo −35.0% more frequent manual overrides. HVAC (summer) $31,000/mo $21,400/mo −31.0% 5) Cross-facility transfer learning accelerated new deploy- Operator hrs/mo 480 h 80 h −83.3% ments.

New facilities with equipment models previously Crop loss/year 3 0 −100% System ROI — 7.2 months — seen in the fleet achieved optimal performance 60–70% faster than early deployments without transfer learning baselines. The continental climate presents distinct challenges: ex- treme winter cold requires managing heating costs while pre- venting condensation; hot humid summers stress dehumidifi- B. What Failed or Required Iteration cation capacity.

The system’s adaptive behavior is particularly 1) Sensor drift at scale was the dominant operational chal- evident during seasonal transitions — automatically adjusting lenge. Across 30+ facilities with thousands of sensors, control strategies as outdoor conditions shift, whereas the sensor drift — not control algorithm performance — was conventional Argus system required manual seasonal recon- the primary source of control degradation.

The median figuration. filtering redundancy (Section II-A) was added after the The 83.3% reduction in monitoring labor reflects the pro- first year of operation when drift-related false alarms gression from L1 to L3 autonomy. Operators shifted from became the leading maintenance issue. continuous monitoring to exception-based supervision. 2) Generic PID auto-tuning was insufficient for some HVAC configurations. The Ziegler–Nichols relay feed- E.

Comparison with Published Literature back method produced acceptable initial gains for ap- Table VIII contextualizes the deployment scale against the proximately 80% of control loops. The remaining 20% existing body of work. — primarily large chilled water systems with significant The difference is not incremental — it represents orders transport delay — required Cohen–Coon identification of magnitude in both facility count and operational duration. or manual refinement.

The neural network adaptation This scale enables observations impossible in short-term de- eventually compensated, but initial performance suffered ployments: long-term equipment degradation detection, multi- during the learning period. season adaptation patterns, and statistically significant cross- 3) Facility operators initially resisted automated CO2 man- facility comparisons. agement. CO2 injection coordination with ventilation cycles required careful cross-subsystem logic.

Early VII. D EPLOYMENT L ESSONS deployments paused CO2 injection during ventilation events but did not account for room volume and air Seven years of production operation across 30+ facilities exchange rate, leading to suboptimal CO2 recovery with 50+ equipment manufacturers have yielded practical times. The system was refined to model room volume lessons that are absent from the simulation-focused literature. dynamics. 4) Legacy equipment integration remained the most time- A.

What Worked consuming commissioning task. While standardized 1) VPD as the master setpoint eliminated cross-coupling BACnet templates worked well for modern HVAC conflicts. The most immediate and consistent improve- equipment, older installations with proprietary serial ment across all facilities was the elimination of simulta- protocols or analog-only interfaces required custom neous heating and dehumidification — a common waste gateway configuration that extended commissioning be- pattern in conventional controllers where independent T yond the 5-day target. and RH loops fight each other. 5) Energy cost coefficients required seasonal recalibration. 2) Edge-first architecture was essential for reliability.

Net- The energy cost parameters (αh , αc , αd , αm ) in the work outages, ISP failures, and cloud service disruptions VPD optimization (4) required quarterly updates to

IEEE ACCESS 8

TABLE VIII

R EAL -W ORLD D EPLOYMENT S CALE C OMPARISON

Study Facilities Duration Domain Control type

All field experiments combined [19] ∼20 buildings 43 days total HVAC Various AI Moshari et al. [21] 13 buildings 138 days District heating RL FarmBeats [22] 2 farms 6 months Open-field ag Monitoring only Cao et al. (iGrow) [11] 1 greenhouse ∼1 season CEA RL Hemming et al. [31] 1 greenhouse 1 season CEA AI-assisted IOGRUCloud (this work) 30+ facilities 7+ years CEA NN-PID + VPD cascade

reflect changing utility tariff structures and seasonal 5) Practical deployment lessons. VPD-based cascade con- equipment efficiency curves. Automating this calibration trol, edge-first architecture, and standardized commis- from utility bill data was added in the third year of sioning are identified as the most impactful design deci- operation. sions; sensor drift management and legacy equipment integration are identified as the dominant operational challenges.

C. Scalability Observations Future work includes federated learning for cross-facility 1) Linear scaling of sensor data, sub-linear scaling of con- knowledge transfer without exchanging raw data, integration trol complexity. Adding zones to a facility scales sensor of dynamic electricity pricing for real-time load optimization, data linearly, but control complexity grows sub-linearly extension to adjacent domains (data center cooling, phar- because adjacent zones share environmental coupling — maceutical cleanrooms), and open-sourcing the standardized the system exploits this coupling for coordinated control.

BACnet equipment templates. 2) Fleet diversity improved model robustness. Counter- intuitively, deploying across diverse climate zones and equipment configurations produced more robust models ACKNOWLEDGMENT than deployments across identical facilities. The diver- The author would like to thank the cultivation teams and sity forced the neural network to learn generalizable facility operators across all deployed sites for their collabora- patterns rather than overfitting to specific equipment tion and operational data.

The author also acknowledges the characteristics. IOGRU engineering team for their contributions to platform 3) 99.9% uptime achieved through redundancy, not re- development and deployment. liability. Individual components failed regularly.

The During the preparation of this manuscript, generative AI 99.9% system uptime target was achieved through edge- tools (Claude, Anthropic) were used for literature search level failsafe modes, sensor redundancy, and graceful assistance and language editing. All technical content, system degradation — not through preventing failures. architecture descriptions, mathematical formulations, experi- mental data, and conclusions are the sole work of the author.

VIII. C ONCLUSION The author reviewed and edited all AI-assisted content and takes full responsibility for the final manuscript. This paper presented IOGRUCloud, a three-tier IoT plat- form for AI-driven climate control in Controlled Environment Agriculture, deployed across 30+ commercial facilities in 8 R EFERENCES U.S. climate zones over 7+ years.

The principal contributions [1] ASHRAE, ASHRAE Handbook — HVAC Applications, Ch. 24: Environ- mental Control for Animals and Plants. Atlanta, GA, USA: ASHRAE, are: 2019. 1) Cascading VPD control. Elevating VPD from a mon- [2] T.

Kozai, G. Niu, and M. Takagaki, Eds., Smart Plant Factory: The Next itored metric to the primary cascading setpoint, with Generation Indoor Vertical Farms.

Singapore: Springer, 2020. [3] N. Engler and M. Krarti, “Review of energy efficiency in controlled envi- neural network-optimized energy-minimal T –RH de- ronment agriculture,” Renew.

Sustain. Energy Rev., vol. 141, p. 110786, composition, achieves 30–38% HVAC energy savings May 2021. and 68–73% VPD stability improvement. [4] E. Iddio, L.

Wang, Y. Thomas, G. McMorrow, and A.

Denzer, “Energy efficient operation and modeling for greenhouses: A literature review,” 2) Neural network PID self-tuning. A 7-3-3 backpropaga- Renew. Sustain.

Energy Rev., vol. 117, p. 109480, Jan. 2020. tion network with Lyapunov stability guarantees adapts [5] R. R. Shamshiri et al., “Advances in greenhouse automation and PID gains online — the first reported production deploy- controlled environment agriculture: A transition to plant factories and urban agriculture,” Int.

J. Agric. Biol.

Eng., vol. 11, no. 1, pp. 1–22, ment in CEA. 2018. 3) Progressive autonomy (L1–L4). A domain-specific adap- [6] R. Sadeghi, H.

Zareiforoush, M. R. Alizadeh, and A.

Jafari, “Advances tation of industrial autonomy concepts enables graduated in intelligent and autonomous greenhouse systems: A comprehensive review,” Smart Agric. Technol., 2025, Art. no. 100901. trust-building with operator guardrails. [7] W.-H. Chen, B.

Decardi-Nelson, C. Kubota, and F. You, “AI applications 4) Scale of validation. 30+ facilities, 930,000+ m2 , 500+ in the environmental control of controlled environment agriculture in the control zones, 7+ years — exceeding the combined du- digital age,” Modern Agric., 2025. [8] A.

Ajagekar, N. S. Mattson, and F.

You, “Energy-efficient AI-based ration of all published AI-based HVAC field experiments control of semi-closed greenhouses leveraging robust optimization in by approximately 60×. deep reinforcement learning,” Adv. Appl. Energy, vol. 9, p. 100119, 2023.

IEEE ACCESS 9

[9] M. A. Adesanya, H.

Obasekore, A. Rabiu et al., “Deep reinforcement ANDRII VAKHNOVSKYI received the M.Sc. de- learning for PID parameter tuning in greenhouse HVAC system energy gree in Computer Engineering and System Pro- optimization,” Expert Syst. Appl., vol. 252, p. 124126, 2024. gramming from the National Technical Univer- [10] I.

Panagopoulos, R. D. McAllister, S. van Mourik, and T.

Keviczky, “A sity “Kharkiv Polytechnic Institute” (NTU “KhPI”), cascaded economic model predictive control approach to greenhouse Kharkiv, Ukraine, in 2011. He is the Founder and climate control,” IFAC-PapersOnLine, vol. 59, no. 23, pp. 443–448, CTO of IOGRU LLC, New York, NY, USA, where 2025. he designed and deployed the IOGRUCloud plat- [11] X. Cao et al., “iGrow: A smart agriculture solution to autonomous form across 30+ commercial CEA facilities in 8 U.S. greenhouse control,” in Proc.

AAAI Conf. Artif. Intell., vol. 36, no. 11, states.

His research interests include industrial IoT, 2022, pp. 11837–11845. adaptive control systems, neural network-augmented [12] S. Zeng, H. Hu, L.

Xu, and G. Li, “Nonlinear adaptive PID control PID control, and energy-efficient automation for for greenhouse environment based on RBF network,” Sensors, vol. 12, controlled environment agriculture. Previously, he worked on building man- no. 5, pp. 5328–5348, 2012. agement systems with Schneider Electric and Siemens authorized integrators [13] A.

Salehi et al., “Robust enhanced auto-tuning of PID controllers for across 20+ industrial projects including data centers, power generation facil- optimal quality control via neural networks,” ChemEngineering, vol. 9, ities, and hospitals. no. 3, p. 52, 2025. [14] T. Gamer, M. Hoernicke, B.

Kloepper, R. Bauer, and A. J.

Isaksson, “The autonomous industrial plant — Future of process engineering, operations and maintenance,” J. Process Control, vol. 88, pp. 101–110, 2020. [15] E. Alreshidi, “A systematic survey on the role of cloud, fog, and edge computing combination in smart agriculture,” Sensors, vol. 21, no. 17, p. 5922, 2021. [16] M.

Sami and K. I. Ibraheem, “Energy-efficient edge-fog-cloud architec- ture for IoT-based smart agriculture environment,” IEEE Access, vol. 9, pp. 110480–110492, 2021. [17] C.

Grossiord et al., “Plant responses to rising vapor pressure deficit,” New Phytologist, vol. 226, pp. 1550–1566, 2020. [18] T. Inoue et al., “Minimizing VPD fluctuations maintains higher stomatal conductance and photosynthesis, resulting in improvement of plant growth in lettuce,” Frontiers Plant Sci., vol. 12, p. 646144, 2021. [19] O. B.

Mulayim et al., “Comparative field deployment of reinforcement learning and model predictive control for residential HVAC,” arXiv preprint, arXiv:2510.01475, 2025. [20] C. Al Sayed, C. E.

Boueri, and N. Youssef, “Reinforcement learning for HVAC control in intelligent buildings: A technical and conceptual review,” J. Build.

Eng., 2024. [21] A. Moshari, K. Javanroodi, and V.

M. Nik, “Real-world deployment of model-free reinforcement learning for energy control in district heating systems,” Appl. Energy, vol. 402, p. 126997, 2025. [22] D.

Vasisht et al., “FarmBeats: An IoT platform for data-driven agricul- ture,” in Proc. USENIX NSDI, 2017, pp. 515–529. [23] R. R.

Shamshiri, J. W. Jones, K.

R. Thorp, D. Ahmad, H.

C. Man, and S. Taheri, “Review of optimum temperature, humidity, and vapour pressure deficit for microclimate evaluation and control in greenhouse cultivation of tomato,” Int.

Agrophysics, vol. 32, no. 2, pp. 287–302, 2018. [24] B. Xu, Y. Fu, G.

Fang, Y. Sun, Z. Du, and Z.

Wang, “One for many: Transfer learning for building HVAC control,” in Proc. 7th ACM Int. Conf. Syst.

Energy-Efficient Buildings, Cities, and Transportation (BuildSys ’20), 2020, pp. 230–239. [25] C. N. Verdouw, B.

Tekinerdogan, A. Beulens, and S. Wolfert, “Digital twins in smart farming,” Agric.

Syst., vol. 189, p. 103046, 2021. [26] O. A. Alduchov and R.

E. Eskridge, “Improved Magnus form approxi- mation of saturation vapor pressure,” J. Appl.

Meteor. Climatol., vol. 35, no. 4, pp. 601–609, 1996. [27] X. C.

Jiao, X. M. Song, D.

L. Zhang, Q. J.

Du, and J. M. Li, “Coordination between vapor pressure deficit and CO2 on the regulation of photosynthesis and productivity in greenhouse tomato production,” Sci.

Rep., vol. 9, p. 8700, 2019. [28] J. G. Ziegler and N.

B. Nichols, “Optimum settings for automatic controllers,” Trans. ASME, vol. 64, pp. 759–768, 1942. [29] M.

Platero-Horcajadas, S. Pardo-Pina, J.-M. Cámara-Zapata, J.

A. Brenes, and F. Ferrández Pastor, “Enhancing greenhouse efficiency: Integrating IoT and reinforcement learning for optimized climate control,” Sensors, vol. 24, no. 24, p. 8109, 2024. [30] F.

Oldewurtel et al., “Use of model predictive control and weather forecasts for energy efficient building climate control,” Energy Build., vol. 45, pp. 15–27, 2012. [31] S. Hemming, F. de Zwart, A. Elings, A.

Petropoulou, and I. Righini, “Cherry tomato production in intelligent greenhouses — Sensors and AI for control of climate, irrigation, crop yield, and quality,” Sensors, vol. 20, no. 22, p. 6430, 2020.

IEEE ACCESS 1

IEEE ACCESS 2

IEEE ACCESS 3

TABLE I

IEEE ACCESS 4

IEEE ACCESS 5

IEEE ACCESS 6

TABLE II

IEEE ACCESS 7

IEEE ACCESS 8

TABLE VIII

J. Agric. Biol.

IEEE ACCESS 9

L. Zhang, Q. J.