How Smart Lighting Cuts Greenhouse Energy Costs by 65% Without Sacrificing Yield

18.976 grams per dollar. That’s the most efficient return on energy spending ever recorded in a commercial greenhouse lighting system — a number so high it nearly doubles the performance of conventional methods. In a side-by-side experiment at Virginia Tech’s Hampton Roads Agricultural Research Center, researchers pitted six different LED lighting strategies against each other in identical conditions, and one emerged as a quiet revolution: a forecast-driven control system that slashes energy costs by 65% while delivering the highest crop yield among adaptive lighting approaches. For an industry where electricity can consume up to 30% of operating costs, this isn’t just incremental progress. It’s the blueprint for a new kind of farming — one where algorithms don’t replace farmers, but empower them to grow more with less.

The Science

The study, led by Sajad Ahmadi, Eric J. Stallknecht, and Javad Mohammadpour Velni (2026), was designed to cut through the noise that plagues agricultural technology research: inconsistent environments, mismatched hardware, and apples-to-oranges comparisons. Instead of testing strategies across different greenhouses or seasons, the team deployed all six lighting control methods — simultaneously — in a single greenhouse at Virginia Tech’s Hampton Roads AREC. Each strategy ran in its own benching zone, with identical lettuce crops (Lactuca sativa L.), shared environmental sensors, and the same LED fixtures. This eliminated confounding variables like weather exposure, shading, or sensor drift, making it the most controlled comparison of its kind.

The platform itself was built on a robust Internet of Things (IoT) architecture using MQTT, a lightweight communication protocol ideal for real-time control in resource-constrained environments. Four Apogee SQ-500-SS quantum sensors measured Photosynthetic Photon Flux Density (PPFD) — the amount of light usable by plants for photosynthesis — every five minutes. These readings were averaged and fed into each control system, ensuring every treatment responded to the same sunlight conditions.

The six strategies tested were:

1. Always-on: LEDs run at full intensity throughout the photoperiod, regardless of sunlight.
2. Always-off: No supplemental lighting; sunlight only.
3. Prediction-based: Uses a Markov chain solar forecast to optimize LED scheduling via a linear program.
4. Heuristic: A rule-based system that calculates the “catch-up” rate needed to meet the Daily Light Integral (DLI) target.
5. Peak-shift (early start): Shifts LED use to off-peak electricity tariff windows, starting at 4:30 AM.
6. Peak-shift (midday start): Similar, but starts at 11:00 AM, minimizing peak power without cost weighting.

The core innovation lies in how the prediction-based and peak-shift strategies use forecasting. The team trained a context-conditioned Markov chain model on seven years of solar data from the National Solar Radiation Database (NSRDB), incorporating season, time of day, and solar position. This model didn’t just predict average sunlight — it captured the probabilistic nature of cloud cover and diurnal patterns, allowing the controller to anticipate when natural light would be sufficient and when LEDs were truly needed.

Figure 1: Greenhouse layout showing benching zones and replicate assignments. Replicates were placed diagonally to minimize shading effects.

What They Found

Over two experimental replicates — from February to April 2026 — the results were striking. The always-on strategy, still common in many commercial greenhouses, delivered a DLI of 17 mol·m⁻²·d⁻¹ (the target) but at a steep cost: $0.2332 per square meter per day. It also produced the highest absolute dry mass at 2.190 grams per plant, but only because it flooded the canopy with light, much of which was redundant or wasted during sunny periods.

In contrast, the prediction-based strategy met the same DLI target at just $0.0807/m²/day — a 65% reduction in energy cost — while achieving a dry mass of 1.531 g/plant, the highest among all adaptive strategies. This means it delivered 80% of the yield of the always-on system for less than a third of the cost.

The heuristic strategy, which requires no forecasting infrastructure, came close — costing $0.0858/m²/day, just a 6% premium over the prediction-based method. However, its yield was lower at 1.302 g/plant, suggesting it delivered light less efficiently, often overlapping with periods of adequate sunlight.

The peak-shift variants achieved the lowest energy costs — $0.0715 to $0.0738/m²/day — by concentrating LED use in off-peak tariff windows. But this came at a biological cost: dry mass dropped to 1.252–1.264 g/plant, a 17–18% reduction compared to the prediction-based approach. This trade-off highlights a critical insight: shifting energy use to cheaper times isn’t always optimal if it means delivering light when plants can’t use it effectively.

Daily Energy Cost and Dry Mass by Lighting Strategy

Label	Value
Always-on	0.2332
Always-off	0
Prediction-based	0.0807
Heuristic	0.0858
Peak-shift (4:30 AM)	0.0715
Peak-shift (11:00 AM)	0.0738

• Always-on 0.2332
• Always-off 0
• Prediction-based 0.0807
• Heuristic 0.0858
• Peak-shift (4:30 AM) 0.0715
• Peak-shift (11:00 AM) 0.0738

Source: Ahmadi et al. (2026). Prediction-based control achieves 65% cost reduction vs. always-on while maintaining high yield.

{{chart:0.title}}

To unify cost and yield into a single metric, the researchers introduced yield-per-dollar efficiency (η$), defined as dry mass per dollar of energy cost. Here, the prediction-based strategy dominated: 18.976 g/$, more than double the always-on baseline of 9.391 g/$. Even the peak-shift strategies, despite their lower costs, scored only 17.1–17.5 g/$,whiletheheuristicreached15.2g/$.

Figure 4: LED commands and cumulative DLI for adaptive strategies on March 3, 2026. The prediction-based controller suppresses LEDs during forecasted sunlight, concentrating output in cheaper, darker windows.

Why This Changes Things

For decades, greenhouse operators have faced a false choice: pay more for higher yields, or save on energy and accept lower productivity. This study dismantles that dichotomy. The prediction-based strategy proves that with the right control logic, you can have both — not by adding more light, but by using it smarter.

The 65% cost reduction isn’t just a number on a spreadsheet. For a 1,000 m² greenhouse, it translates to saving over $50,000 annually in electricity alone. At scale, such savings could make indoor farming viable in regions where energy costs have historically been prohibitive.

But the real breakthrough is the yield-per-dollar metric. Traditional agricultural economics focuses on yield per unit area or energy per unit yield. By combining both into a single efficiency measure, the study reframes the goal of greenhouse management: it’s not just about growing more, but about growing more profitably. This shifts the conversation from input maximization to return optimization — a mindset aligned with sustainable intensification.

The heuristic strategy’s near-parity with the prediction-based method is equally significant. It suggests that even without advanced forecasting, simple rule-based systems can achieve most of the savings. For small-scale or resource-limited growers, this is empowering: you don’t need a PhD in machine learning to cut your energy bill by 60%.

Moreover, the study exposes the limits of peak-shifting. While demand charges and time-of-use tariffs incentivize shifting energy use, doing so without regard for plant physiology can backfire. The peak-shift variants saved money but sacrificed yield, revealing that light timing matters as much as light quantity. Plants don’t just need photons — they need them at the right time, when their photosynthetic machinery is primed to use them.

This has implications beyond lettuce. As climate change disrupts growing seasons and urban agriculture expands, the ability to precisely match light supply with plant demand will become a cornerstone of food security. The Markov chain forecasting model, though simple, demonstrates that even modest predictive tools can outperform static rules — a lesson applicable to irrigation, heating, and ventilation systems.

What’s Next

The study opens several paths forward. First, the prediction-based controller used a linear program (LP) with a fixed horizon and re-planned every five minutes. Future work could explore model predictive control (MPC) with longer horizons or adaptive learning, potentially closing the gap between forecast-based and perfect-knowledge performance (

shows this gap persists, indicating room for improvement).

Second, the yield-per-dollar metric could be expanded to include labor, capital, and carbon costs, creating a true profit-per-resource index. As carbon pricing gains traction, such metrics will be essential for comparing the full environmental and economic footprint of different farming systems.

Third, the heuristic strategy’s 6% cost premium over prediction-based control raises a practical question: is the added complexity of forecasting worth it? For large commercial operations, likely yes. But for smaller farms, the answer may be no — suggesting a tiered adoption pathway where simplicity precedes sophistication.

Finally, the study was conducted in Virginia during late winter and early spring, when sunlight is limited. Testing these strategies in sunnier climates or during summer months could reveal different trade-offs. In regions with abundant sunlight, the always-off or minimal-supplement strategies might dominate, further shifting the cost-yield frontier.

One limitation is clear: the model assumes a fixed canopy absorption limit ($\bar{P}$) and doesn’t account for dynamic plant responses like photoinhibition or circadian rhythms. Future controllers could integrate real-time plant feedback — via chlorophyll fluorescence or thermal imaging — to adjust light delivery based on actual physiological need, not just predicted availability.

The broader lesson is that the future of agriculture isn’t just about bigger greenhouses or brighter LEDs. It’s about smarter decisions. As Ahmadi and colleagues have shown, a well-designed algorithm can turn a fixed cost into a dynamic asset — one that responds to weather, markets, and biology in real time. This isn’t automation replacing farmers; it’s augmentation, giving growers a new lens through which to see their crops, their energy use, and their bottom line.

In a world where food systems must produce more with less land, water, and energy, such precision isn’t a luxury. It’s a necessity. And the path forward may not be in the soil, but in the code.