Direct reinforcement-learning control of refrigeration systems across a distributed grocery fleet.
Koios runs PPO and DQN policies that write the expansion valve and compressor stage directly, learning each case's superheat and duty-cycle behavior instead of holding the conservative fixed setpoints legacy controllers default to. A confidence model gates every move; whenever the agent is uncertain or out of its trained envelope, the store's existing Danfoss or Copeland controller takes the loop back, colder food and lower energy with the vendor system always underneath.
A supermarket runs dozens of refrigerated cases and walk-ins, each metered by an expansion valve and served by a compressor rack. Legacy Danfoss, Copeland/Emerson E2, Carel and Honeywell controllers hold fixed superheat and suction-pressure setpoints chosen for worst-case safety, so cases run colder and compressors cycle harder than the load actually requires, wasting energy, accelerating wear, and still drifting into temperature excursions when conditions swing.
Koios deploys deep-reinforcement-learning agents that learn each circuit's response and actuate the electronic expansion valve and compressor staging directly, over a bi-directional, secured comms bridge to the store's existing controllers. A PPO policy trims continuous superheat and valve position; a DQN policy makes the discrete calls, staging and valve steps, and an AI Control Confidence model hands the loop back to the native controller outside the trained envelope. The same software stack runs vendor-agnostically across 2,585 stores and 1,300 Koios systems in nine regions, even on antiquated controllers, so Koios dominance and RL control prevail across the estate.
Ai-OPs
·Direct EEV RL Control01
Why grocery refrigeration leaves savings on the table
Refrigeration is the single largest electrical load in a supermarket, and almost every case is run on a fixed, conservative setpoint that the load rarely justifies.
Each refrigerated case and walk-in is fed through an expansion valve that meters liquid refrigerant into the evaporator and controls superheat, the margin that keeps liquid from reaching the compressor. Legacy controllers, thermostatic valves or electronic valves on a fixed superheat target, are tuned for the worst case: high superheat so the compressor is never flooded, low suction pressure so the warmest case still holds. Safe, but structurally inefficient.
Where the waste and risk come from
- Over-cold cases: a fixed setpoint holds product colder than the food-safe band requires whenever ambient and door load are below worst-case, burning energy for no benefit.
- Starved evaporators: conservative superheat keeps a slug of the coil dry, cutting heat transfer so the compressor runs longer to do the same cooling.
- Short-cycling compressors: fixed suction floats and on/off staging cycle racks far more than the load needs, the dominant driver of wear, oil issues and call-outs.
- Excursions anyway: static setpoints can't anticipate defrost, door surges or rack interactions, so product temperature still drifts out of band during transients.
A fixed setpoint has one operating point. Refrigeration load moves continuously with ambient, traffic, defrost and product. A reinforcement-learning policy learns that response and acts on the live trend, tightening superheat and easing the compressor exactly as far as conditions allow, then backing off before food safety is at risk.
FIG. 1: THE CONTROL LEVER. Refrigerant flows compressor → expansion valve → evaporator → back to suction. The valve sets superheat, which couples evaporator efficiency, case temperature and compressor duty. A fixed setpoint fixes all three at a worst-case compromise.
Ai-OPs
·Direct EEV RL Control02
Closed-loop control architecture
A Koios Gateway Manager bridges the store's controllers to Koios over secured, bi-directional comms. Koios reads every case tag, the RL agent writes the valve and stage directly, and a confidence model keeps the native controller as a live fallback.
MANAGER
BRIDGE
FIG. 2: VENDOR-AGNOSTIC CONTROL & CONNECTIVITY. The Gateway Manager normalizes every controller family into one bi-directional, secured comms session. Koios writes the expansion valve and compressor stage directly; the confidence model returns the loop to the native controller whenever inputs leave the trained range.
What the agent reads
Per circuit: case / product temperature, superheat, suction and discharge pressure, valve position, defrost state, ambient and compressor duty. Koios manages historian quality and range tolerances, then inferences each control scan on a rolling window so the policy acts on a trend, not a single noisy sample.
What it writes
Directly: EEV position / superheat target per case and suction float and staging per rack, positional commands clipped to safe ranges. The agent acts only while every input stays inside its trained envelope; outside it, the store's existing Danfoss or Copeland controller holds the loop, so a poorly-characterized state never reaches the valve through the model.
Ai-OPs
·Direct EEV RL Control03
Templatized control models, deployed at hyperscale
Koios doesn't hand-build a controller for each store. It maintains a catalog of templatized control models, proven, reusable model archetypes the system selects, deploys and tunes per site automatically. PPO and DQN are two such templates, applied where each fits; the platform, not an engineer, runs them.
A Proximal Policy Optimization template for problems where a setpoint is trimmed smoothly and continuously, EEV position, superheat target, suction-pressure float. Reused across product lines wherever fine, always-on regulation beats a fixed setpoint.
A Deep Q-Network template for problems that are a choice among modes or steps, compressor staging, sequencing and coarse valve steps. The system applies it wherever the decision is discrete, and it travels to other product lines unchanged.
What the templates optimize for
Each template is trained against a shared objective that encodes the operator's real trade-offs, biased, like the field configuration, toward food safety and calm equipment motion.
FIG. 3: REWARD STRUCTURE. Energy, equipment wear and food safety pull against each other; the configuration is biased toward safety and smooth action, with the confidence guardrail as a hard backstop.
A separate AI Control Confidence model scores every proposed action against the conditions the model was trained on. High confidence, in-range: Koios writes the valve. Low confidence, novel state, or a stale tag: control reverts bumplessly to the store's native float. The vendor controller is never removed; it sits underneath as a continuous, automatic fallback, so the model never owns the valve outright.
Ai-OPs
·Direct EEV RL Control04
Measured results in a live store
Against a weather-normalized July baseline regression, the AI-controlled racks ran a consistently lower duty cycle through October; the shaded efficiency cloud below the fit line is realized energy the compressors no longer spend.
FIG. 4: BEFORE. Observed vs. predicted duty cycle scatters evenly around the 1:1 line; the baseline simply fits itself, no systematic efficiency gain.
FIG. 5: AFTER. Observed duty falls well below what the frozen baseline predicts; the points lift into a "better-efficiency cloud" above the line. That gap is 26%+ energy savings.
FIG. 6: DUTY-CYCLE TRACE. Under fixed setpoints the rack cycles constantly; under RL control the same load is met with far fewer, smoother compressor events, the source of the 80% cycle reduction and the maintenance dividend.
Energy figures are duty-cycle savings against a weather-normalized baseline (IPMVP-style M&V). Reported values are demonstrable averages; actual results vary by store, equipment family and climate, and maintenance savings are estimates based on the measured reduction in equipment movement.
Ai-OPs
·Direct EEV RL Control05
Distributed control at fleet scale
The same Koios software stack runs every store. One Koios per up to two locations, grouped into nine cloud regions, onboards identically, antiquated or modern, Danfoss or Copeland, so RL control reaches the whole estate without a per-store rebuild.
Why dominance holds, even on legacy gear
The Gateway Manager bridges every controller family into one secured comms session, so a store on a decade-old Danfoss AK-SC or a Copeland E2 onboards through the same path as a new SM-800 site. Bi-directional data flow means Koios both reads the estate and writes control back, and the central Model Deployment API pushes the equipment-matched RL package automatically. No other system consolidates a mixed, multi-vendor estate and executes control across it the way the Ai-OPs Koios platform does.
- Same Koios software stack on every store, vendor-agnostic and case-style agnostic.
- Handles 350+ models and 10,000+ data points per integrated site without manual configuration.
- Read-only fleet monitoring in Koios Lens, KPIs, alarms and diagnostics, architecturally separated from control.
- Tested across 48 refrigeration types, legacy to modern, fully validated.
- Proven across continents, production deployments in the US, UK, Europe, Australia and Southeast Asia.
| Region | Stores | Koios | Models |
|---|---|---|---|
| Region 01 | 699 | 350 | 27,963 |
| Region 02 | 348 | 174 | 13,921 |
| Region 03 | 344 | 174 | 13,761 |
| Region 04 | 268 | 135 | 10,721 |
| Region 05 | 257 | 129 | 10,281 |
| Region 06 | 214 | 108 | 8,561 |
| Region 07 | 208 | 105 | 8,321 |
| Region 08 | 157 | 79 | 6,281 |
| Region 09 | 90 | 46 | 3,600 |
| 9 regions | 2,585 | 1,300 | 103,410 |
TABLE 1: TYPICAL DISTRIBUTION. A representative regional rollup for a larger grocer; the RL control footprint scales linearly as regions come online.
Read/write RL control stays at the store Koios for resilience; a regional or cloud disruption never strands a valve. Koios Lens consolidates the fleet read-only, so enterprise monitoring and Curtis GPT context can never actuate equipment. Monitoring and control are separated by design.
Ai-OPs
·Direct EEV RL Control06
Where this applies
The same pattern, RL policies that learn a circuit and write the valve directly, with the native controller always underneath, generalizes across the refrigerated estate. Available today on Koios:
About Koios
Koios is Ai-OPs' self-hosted machine-learning inferencing platform for deploying scalable AI models on industrial and refrigeration control systems. It is hardware- and vendor-agnostic, runs on-premises or distributed in the cloud, and writes optimized control back to PLCs, SCADA, DCS and refrigeration system managers in real time. Your models and data stay inside your estate; you own every model Koios trains.
It installs as a single Docker image, paired with the Koios Gateway Manager bridge per store and consolidated read-only in Koios Lens. Licensing is an annual per-instance subscription under a separate commercial agreement.
| Control families | Danfoss · Copeland/Emerson E2 · Carel · Honeywell · Schneider |
| Protocols | Danfoss XML · BACnet · Modbus TCP · REST · secure comms & more |
| RL policies | PPO (continuous) · DQN (discrete) · confidence-gated |
| Deployment | Per-store Koios · 1 per up to 2 sites · regional cloud |
| Fallback | Native controller, bumpless & automatic |