Data centres: cooling on worst-case settings, power for nothing

Cooling runs on worst-case settings

What happens. Chilled-water temperature, supply air temperature and fan speeds are set for the hottest day and the fullest hall, then left alone. Most of the year the load is lower and the outside air is cooler, but the setpoints stay where they were set at commissioning.

What it costs. Power for cooling that the servers did not need. Free cooling is switched in late or not at all, and chillers run in part load, where they are least efficient.

Why the usual fix fails. Raising setpoints by hand saves power until a hot spot appears, and then the margin returns. Operators are judged on availability, not on the power bill, so they keep the safe setting.

How a fixed-price remote diagnosis checks it. From your building management and power logs (temperatures, valve and fan positions, chiller states, server load, outside conditions) I check how much of the cooling energy goes to margin that was never used. The fixed-price remote diagnosis tells you whether predictive cooling, which acts ahead of load and weather, can hold your limits with less power, and what it would need.

Background: Economic predictive control

Cooling plant control is tuned on site, without a physical model

What happens. Chillers, pumps, cooling towers and heat exchangers are controlled by rules and tuned on the plant, one change at a time. Nobody has a simulation of how the whole cooling circuit behaves, so a new control idea can only be tried live, in a building where nothing may go wrong.

What it costs. Slow commissioning, cautious settings and years of living with a control that is only good enough. Every extension needs the tuning redone by someone who knows the plant.

Why the usual fix fails. A purely data-driven model needs data from situations that were never run, such as extreme loads or partial failures. A hand-built simulation by the plant builder is often not kept up to date and does not match the plant as it is.

How a fixed-price remote diagnosis checks it. From the piping and instrument drawings, equipment data and a few weeks of logs I check how much of the plant can be described by a physical model that is calibrated on your data. The fixed-price remote diagnosis tells you whether such a model is within reach and whether it could be used to design and test the control before it runs live.

Background: Predictive control of process plants

Efficiency and heat-reuse rules meet chillers on fixed setpoints

What happens. Rules for data centres now set targets for energy efficiency, for reuse of waste heat and for the source of the power. Chillers, free cooling and heat recovery each run on their own fixed setpoints and sometimes work against one another, for example when heat recovery wants warm return water and the chillers want cold.

What it costs. The efficiency figure that the rules measure is worse than the hardware could deliver, and heat that is meant to be reused is lost. Missing a target can be costly, and so can over-building equipment to be sure.

Why the usual fix fails. More equipment, such as bigger heat exchangers or extra chillers, adds investment but not coordination. Fine-tuning each loop by hand does not help when the loops interact.

How a fixed-price remote diagnosis checks it. From your metering, cooling and heat-recovery logs and the targets you must meet, I check where the plant loses efficiency and whether coordinated control could reach the targets with the equipment you have. The fixed-price remote diagnosis tells you whether economic predictive control of the whole cooling plant, including heat-recovery scheduling, is worth building. Rafał has worked on predictive control of the superfluid-helium cooling of the Large Hadron Collider at CERN, the European particle-physics laboratory, a cooling plant with similar coupled loops.

Background: Economic predictive control

Server waste heat is too cool for the heating network

What happens. A data centre gives off a great deal of heat, but at a low temperature. Air-cooled halls deliver lukewarm air, and even liquid cooling reaches less than a district heating network normally needs. To sell or hand over the heat, a heat pump has to lift it, and the server load follows the customers, not the heating demand.

What it costs. Heat pump energy, and heat that is wasted in summer or at night when the network has little use for it. If the data centre and the network are run separately, each pays for the other’s mismatch.

Why the usual fix fails. Fixing a supply temperature in a contract is simple but ignores that load and demand both swing. A larger storage tank helps, but it costs money and does not tell when to charge it.

How a fixed-price remote diagnosis checks it. From your cooling and load logs, and the temperature and demand profile of the network if you have it, I check how much heat is available at which temperature and when. The fixed-price remote diagnosis tells you whether joint control of cooling, heat pumps and network supply could raise the usable share of heat, and what data both sides would have to share.

Background: Economic predictive control