Sensors and machine-learning models are getting good at forecasting loss points days to weeks ahead—turning “reactive” storage into predictive control. The prize: fewer downgrades, tighter energy spend, and smarter interventions before quality slides.
Key points:
- Data stack – tuber temperature, CO₂/O₂, relative humidity (RH), airflow/pressure, selective gas traces, and defect imaging feeding supervised models.
- Actionability – alerts must translate into fan speed changes, set-point nudges, sprout-control timing, and pull-date shifts, not dashboards for their own sake.
- Quality focus – forecasting reducing sugars (for fryer color), pressure-bruise risk around handling/early dehydration, and likely sprout break windows.
- Integration – linking store data to field histories (kill date, soil moisture, variety, handling) to set realistic baselines and thresholds.
- ROI – pilots and extension case studies indicate lower shrink (≈0.4–1.7% in VFD fan studies) and 30–45% fan energy savings; benefits are constrained by legacy sheds with limited controllability.
Why predictive storage is coming into focus
Most storages already log temperature and RH. Many now watch CO₂/O₂ and run variable-speed fans. What changes with AI is the ability to fuse those time-series signals with context – cultivar, harvest maturity, field moisture – and learn signatures that precede quality failures.
In grain, for instance, researchers have shown that temperature, RH and CO₂ streams can be used with machine learning to predict storage quality loss risk – a strong methodological analog for potatoes.
In cellars, the same principle applies: use continuous sensing to flag incipient problem zones (a rising respiration pocket, a developing moisture gradient, a subtle airflow imbalance) and trigger small, timely adjustments that avoid bigger losses later.
Building the data stack that models can learn from
Environmental signals. Core or near-core tuber temperature, ambient RH, CO₂ (and sometimes O₂) form the base layer. Well-sealed stores are managed to ≈90–100% RH to minimize weight loss, with airflow carefully balanced to avoid cold spots and condensation. Ventilation design – bulk vs. box, positive vs. space ventilation – matters for how signals evolve and how control works.
Airflow and pressure. Differential pressure across plenums and ducts helps map how air reaches (or fails to reach) zones in the pile. Computational work from Wageningen has modeled pressure flattening under bulk storage, showing why persistent load plus local dehydration can set up damage long before you see it.
Gas traces and volatiles. Beyond CO₂, research and prototypes are using volatile detection – including electronic noses – to pick up early soft rot and other disorders. Both lab and store trials show pre-symptomatic detection is feasible, which is ideal as a predictive feature.
Imaging/defect sensing. Non-destructive approaches (VIS/NIR, hyperspectral, ultrasound, X-ray/CT) are increasingly accurate for detecting internal and external defects. These tools help create labels for model training and can also serve live in-store where feasible.
Operational metadata. Fan speeds, damper positions, vent cycles, humidification, and set-point histories are essential. They connect cause and effect – the levers the model can recommend moving.
Ground-truth labels. Reliable training needs outcome data: bruise incidence by zone or lot, sugar profiles and fryer color, sprout break timing. Without this, models will be brittle. Post-storage assays tied to sensor timelines are the limiting input in many operations.
From signal to decision: alerts that actually change outcomes
Alerts only matter if they map to control. Good systems do four things:
- Prioritize by risk and time-to-action, so managers see what needs intervention now vs later.
- Translate into specific levers – e.g., “lower VFD fan speed 5% for 8 hours in zones 2–3; target RH +0.5%,” or “advance ethylene/mint oil cycle by 48 hours.”
- Show the “why” – a short explanation (CO₂ trend + RH dip + pressure gradient) to build trust.
- Close the loop – capture the response and check whether risk subsides, improving the model next time.
What to predict – the big three loss drivers
1) Pressure bruise (and its precursor, pressure flattening)
Pressure bruise shows up as flattened or depressed areas that develop in storage, typically from dehydration and static load. Extension guidance has emphasized that low RH early in storage increases dehydration and bruise susceptibility; conditioning with high RH is standard to let wounds heal and limit water loss.
Store design matters. Deep bulk piles are efficient, but very deep stacks can increase compression risk; UK guidance warns of pressure bruising at bulk depths over ~4 m, especially without robust positive ventilation and tight humidity control.
Predictive control models can watch depth-related pressure proxies and microclimate signals to flag high-risk zones before visible symptoms. To avoid flattening progressing to bruise, predictive nudges typically focus on slowing local airflow that is desiccating a face, raising RH modestly, and smoothing temperature differentials that cause condensation or rehumidification cycles.
Teagasc also notes that uneven temperatures drive condensation and softening – both implicated in damage risk.
2) Sugar ends and fryer color shifts (cold-induced sweetening)
Cold-induced sweetening (CIS) is well defined: storing tubers cold drives starch breakdown to reducing sugars (glucose, fructose), darkening fries and raising acrylamide risk. Multiple reviews and new work confirm the physiology and the storage-temperature link.
Predictive features include small changes in respiration or CO₂, sub-degree temperature drift, and lot-specific maturity or nitrogen history that predispose certain cultivars. Because sprout suppression regimes also interact with sugars, models should incorporate the chemistry of chosen inhibitors.
Continuous ethylene can suppress sprouting but may increase reducing sugars in some cultivars – a trade-off to navigate before fryer color drifts. Predicting sugar endpoints lets managers adjust temperatures, venting cadence, or sprout-control strategy sooner.
3) Sprout break windows
Dormancy length is a function of cultivar, maturity, field conditions, and store environment. Modern programmes integrate maleic hydrazide (MH) in the field with in-store tools such as ethylene, mint or orange oils, DMN or 3-decen-2-one depending on jurisdiction. Predicting when a lot is about to break dormancy (from subtle respiration or volatile signals) allows timely dosing or a pull-date shift without overuse.
Why field history belongs in the cellar model
Lot-level context makes predictions sharper. Kill date, pre-harvest soil moisture, nitrogen rate, harvest temperature, handling drops, and cultivar all change storage behavior. Classic references from CIP and FAO and contemporary reviews underline the pre-storage determinants of later losses – which is exactly the metadata an AI system needs for a solid baseline.
A practical setup links harvest logs to storage IDs at intake. From day one, the model “knows” that a slightly immature lot of Ranger Russet with low soil moisture at harvest needs stricter dehydration control and a more conservative airflow profile.
What the ROI looks like – and the constraints you’ll hit
Where there are controllable levers (VFD fans, zoned ventilation, humidification, reliable valves or dampers), predictive systems can deliver:
- Energy savings. Extension case studies in potato storage show variable-speed drives cutting fan energy roughly one-third to nearly one-half, depending on season length and design.
- Lower shrink. Side-by-side Idaho data reported 0.4–1.7% lower shrink with VFDs vs constant-speed fans, reflecting better humidity and airflow management – savings that predictive control should at least match or improve.
- Fewer downgrades. The main value is avoided fryer-color penalties and bruise claims. Hard numbers vary by factory and season, but the mechanism is straightforward: anticipate CIS risk and dehydration pockets early and correct gently.
Constraints are real:
- Legacy sheds. If zones can’t be independently ventilated or fans are fixed-speed, models have limited power to help.
- Sensor fidelity. Low-cost probes drift; poor calibration will erode model accuracy over time.
- Label scarcity. Many operations don’t routinely tie cut or fry tests and bruise grading back to time and location in the pile – starving the model of ground truth.
- Change management. Managers will ignore black-box alerts that don’t explain themselves. Strong “why” text and clear confidence ranges help adoption.
A practical, staged adoption path
- Instrument one bay properly. Add dense temperature/RH, CO₂, and pressure points; if possible include a small volatile sensor test. Validate sealing and humidification so the model sees clean signals.
- Digitize field context at intake. Capture kill date, moisture, cultivar, and handling notes as structured data tied to lot ID.
- Start with rule-based alerts, then add ML. Begin with extension-derived thresholds for dehydration and CIS risk; layer a supervised model once you have one full season of labeled outcomes.
- Connect alerts to levers. Ensure recommendations specify fan %, RH set-point tweaks, or sprout-control timing – and log the outcome.
- Run a parallel trial for one season. Compare predictive-assisted control vs business-as-usual on matched lots for energy, shrink, and quality claims.
What’s coming next
Three developments are likely to mature fast:
- Volatile sensing at scale. Electronic noses and targeted gas arrays are moving from lab to store, enabling pre-symptomatic detection of soft rot and possibly signaling dormancy shifts.
- Better non-destructive quality scans. VIS/NIR and hyperspectral tools are getting cheaper and more robust; they’ll increasingly supply live labels and spot checks inside storage and in-line at processing.
- Digital twins of storages. CFD models combined with live sensors will let systems “try” a control move virtually before committing – a safer way to correct emerging hot or cold spots.
Bottom line
Predictive storage isn’t about replacing the operator’s eye; it’s about hearing the store earlier. The combination of well-understood postharvest science (humidity, temperature, airflow), stronger sensing (including volatiles), and pragmatic machine learning offers a credible path to fewer bruises, steadier fryer color, and calmer sprout control – all while trimming kWh per tonne. The trick is to make the signal actionable, wire it to real levers, and keep the model honest with ground-truth tests.
Auteur : Lukie Pieterse, La pomme de terre à l'état pur
English 
Português do Brasil 
Français du Canada 
Español de México