Skip to content
DYCOMAUTOMATION
← Field notes

Cooling on a forecast — why a chiller that never reaches setpoint is costing you money

A cold plunge with a 4°C target that had never been below 5.2°C, and a chiller running eighteen hours a day to get nowhere. A real build: what a year of logged data said about the plant, the setpoint and the thermal fabric — and why the inefficient window turned out to be the cheap one.

Dycom Automation Thermal control Home Assistant Energy Solar

A chiller that never reaches its setpoint is not a chiller. It is a continuously running load with a thermostat attached for decoration, and it will quietly cost you more every month than the equipment did.

This is a write-up of a real build: a cold plunge with a small chiller, a tariff worth optimising against and a target temperature it had never once achieved. The interesting part is that the equipment turned out to be fine. The control strategy, the thermal fabric and the setpoint were the problem — which is almost always the case, and is why we ask for a month of data before we quote a fix.

There is a longer and considerably blunter version of this story — the wrong turns, the swearing and the full configuration — on Angry Dad Automates, the community side of what we do. That write-up has every line of YAML and the fitted model. This one has the reasoning, the costs and the bits a client actually needs to decide on.

The symptom, and why it was misread

The complaint was simple: the water was supposed to be at 4 °C by 5:30am and was sitting closer to 12 °C. The obvious reading is that the chiller is undersized or failing, and the obvious quote is a bigger chiller.

That would have been the wrong job. A year of logged data — water temperature, chiller power, outside air and solar generation, all retained in a long-term database rather than Home Assistant’s default ten days — told a different story.

What a year of data showed

  • The plant was smaller than the control logic assumed. The chiller draws about 300 W, steady, 287–314 W across hundreds of hours. The solar-surplus trigger had been configured to wait for roughly 700 W of export before starting, so for a week it had been declining free energy it could have used twice over. Metering the load before writing the threshold would have caught it on day one.
  • Runtime was already at the ceiling. In December the chiller ran 566 hours in a month — eighteen hours a day. It was not under-utilised. It was saturated.
  • The vessel leaks heat at a measurable rate. Regression against a year of readings gives roughly 0.03 °C per hour for every degree of air-to-water difference. Hold 8 °C water on a 28 °C afternoon and that is a 20-degree gradient, so about 0.6 °C an hour of heat gain — which is approximately the entire output of the machine. At that point it is not cooling. It is holding.
  • Afternoon solar gain made it worse. The vessel sits under a patio roof, but low western sun reaches it in the afternoon and it is black. Heat gain per degree of gradient ran about a third higher in the afternoon than in the morning for the same air temperature.
  • The cheap window has a hard limit. The super off-peak rate runs midnight to 6am. The chiller pulls roughly a degree an hour, slowing as the water cools, so the overnight run is worth about 6 degrees total — no matter how the schedule is written.

Set a target the plant can actually reach

From the fitted model, the coldest the water can get with the chiller running continuously, by outside air temperature:

  • 20 °C air — about 5 °C water
  • 24 °C air — about 8 °C water
  • 28 °C air — about 11 °C water
  • 32 °C air — about 13.5 °C water

The recorded minimum across the whole year is 5.2 °C. The setpoint was 4 °C. In South-East Queensland that is not a setpoint, it is an aspiration.

This matters far beyond the comfort of the person getting in. An unreachable setpoint means the controller is never satisfied, so it calls for cooling every minute it is permitted to — including the hours when it is achieving almost nothing. On one 35 °C day this installation spent about a kilowatt-hour between 11am and 2pm to move the water 0.2 of a degree. The setpoint was moved to 8 °C: achievable, still cold enough to be unpleasant, and crucially a number the thermostat can reach and switch off at.

Efficiency and cost are not the same question

This is the part that usually gets decided on intuition and usually gets decided wrong.

A chiller is a heat pump in reverse: it performs best with warm water and cool ambient air, and worst with cold water and hot air. Per kilowatt-hour, this one does roughly three times as much cooling at 2am as it does at 2pm. The engineering instinct is therefore to block afternoon operation entirely and let the cheap, efficient overnight window do the work.

Costing it changes the answer. Afternoon operation runs on surplus solar that would otherwise be exported at a feed-in tariff of about 2c/kWh. Overnight operation buys grid energy at about 8c/kWh. Run the numbers per degree of cooling delivered:

  • Afternoon, on surplus solar, at its worst efficiency: about 1.5c per degree
  • Overnight, on super off-peak grid, at its best efficiency: about 1.8c per degree

The inefficient window is the cheaper one, because the opportunity cost of self-consumed solar is almost nothing. And since the overnight run is capped at about six degrees regardless, every degree removed during the day is a degree the night does not have to find. Blocking afternoon operation would have produced a warmer result at 5:30am in order to save a few cents.

Optimise against the tariff, not against the coefficient of performance. They point in opposite directions more often than you would expect, and only one of them appears on the bill.

Forecast the result, do not react to the temperature

A thermostat answers “is it cold now”. Nobody cares. The question that matters is “will it be cold at the time it is needed”, and that is a forecasting problem.

The model is deliberately simple and fitted entirely from site data. With the chiller running, the water decays exponentially towards a floor temperature that is a linear function of overnight air temperature, closing roughly 10% of the remaining gap each hour. With the chiller off, it drifts upward towards ambient at the measured gain rate. Overnight air comes from the Bureau of Meteorology forecast, because planning for tonight requires knowing what tonight is going to do.

Validated against a week of real nights, it predicted the 5:30am water temperature to within about half a degree.

That turns one unanswerable question into three numbers on the wall tablet: the projected temperature at the ready-by time, the warmest the water can be at midnight and still make it, and where it is actually heading by midnight. On the morning this was commissioned, those read 10.9 °C projected, 12.2 °C required, 17.2 °C projected — an unambiguous “not on track” eight hours before the failure would have been discovered the hard way.

And a fallback window when it will not make it

If the projection says midnight will arrive too warm for the overnight run to recover, there is one more window worth buying: 10pm to midnight at the off-peak rate of about 25c/kWh. More expensive per kilowatt-hour, but it sits immediately before the cheap run, so very little of it leaks back out before it counts.

Worst case it adds about 15c a night and buys a degree and a bit at 5:30am. It is exposed as a switch, it defaults to off, and it only ever engages when the forecast says the cheap path will fall short. An optional spend the owner has opted into is a very different thing from a controller quietly deciding to buy peak energy.

Protect the compressor from the control logic

Commissioning on an overcast day exposed a defect that a clear day would have hidden for months. With solar output fluctuating behind cloud, the surplus-following logic was stopping the chiller and restarting it two to six minutes later, repeatedly.

Short-cycling is how compressors die. They need several minutes off for head and suction pressures to equalise before restarting, and every start draws heavily while the motor comes up to speed. A minimum cycle time had been specified — but it had only ever been enforced on the stop transition. Restarting was unconstrained.

The minimum off-time now gates the start condition as well. It is a one-line fix and it is the difference between a compressor that lasts its service life and one that does not.

Any load that modulates against solar needs anti-short-cycle protection on both edges, tested under the conditions that actually trigger it. Pumps, compressors and contactors all pay for this one.

Fix the thermal fabric before tuning the controls

Everything above is control strategy, and control strategy cannot beat physics. The model is blunt about the hierarchy: halve the heat gain into the vessel and the overnight run alone reaches 8 °C by 5:30am, on about 13c of energy a night. That is less power than the current arrangement and a colder result, achieved with no additional plant whatsoever.

On a client site the recommendations fall out directly:

  • An insulated cover with a light-coloured upper surface. A thin dark lid on a dark vessel in direct sun is a solar collector. This is the single highest-return item on the list.
  • Insulate the base. Even with zero air-to-water gradient the vessel still gains about 0.1 °C an hour, which points at conduction from the slab through a thin floor. Closed-cell foam underneath is inexpensive and permanent.
  • Shade the vessel from low afternoon sun, without obstructing the plant.
  • Verify by metering, not by impression. The pre-change steady draw is on record at 287–314 W. If the same machine starts sitting near 350 W on comparable days, it is re-breathing its own exhaust, and the data will say so before anyone notices the water is warmer.

Respect the condenser airflow. This is where the interim shade arrangement on this site went wrong: the equipment sits in a corner bounded by a wall, a glazed door and the vessel, and the shade closed off the last open face. A chiller rejecting heat into a pocket it cannot clear will re-ingest its own exhaust and lose efficiency — and if that warm air washes across the vessel wall, the shade is now a heater. Maintain the manufacturer’s clearances, and verify after installation rather than assuming.

A pop-up shade shelter erected over a black inflatable cold plunge on a residential patio, with the chiller enclosed in the corner behind it.
The interim shade arrangement, and the problem with it. Effective against low afternoon sun — but the vessel and the plant share a corner bounded by a wall and a glazed door, and the shelter has closed off the remaining open face. Shade the vessel, not the condenser.

A necessary note on scope and safety. This write-up covers monitoring and control logic. Fixed wiring, dedicated circuits, outdoor outlets and anything connected to the plumbing are work for the appropriately licensed trades, and equipment must be installed to the manufacturer’s instructions and the relevant standards. Automation changes when plant runs; it does not change what a circuit is rated to carry, and it must never be the only thing preventing an unsafe condition. Cold-water immersion also carries genuine health risks for some people — that is a conversation for a doctor, not a control system. We are glad to do the automation on top of a compliant installation; we will not do it instead of one.

What it is worth

No additional plant was purchased. The changes were a corrected surplus threshold, an achievable setpoint, a forecast-driven schedule with an opt-in fallback window, proper anti-short-cycle protection, and a list of fabric improvements ranked by measured return.

The outcomes that matter to an owner: the equipment now reaches its target and switches off, instead of running eighteen hours a day towards a number it could never hit. Energy is bought in the cheapest window that will do the job, and the system says in advance when it will not be enough rather than reporting the failure at 5:30am. The compressor is no longer being short-cycled by passing cloud. And the remaining gains are quantified, so the next dollar spent is spent on insulation — which the data says returns more than any upgrade to the plant.

The same approach applies well beyond a cold plunge. Pool heating, spa plant, server-room cooling and refrigeration all share the pattern: a setpoint nobody validated, a schedule that ignores the tariff, and a thermal envelope nobody measured.

The full configuration

The complete Home Assistant package — the fitted thermal model as a reusable template, the forecast sensor, the scheduling logic and the cycle protection — is published in full in the Angry Dad Automates write-up linked at the top. It is there to be copied.

If you would rather not spend a month fitting exponential decay curves to your own backyard, that is what we are for.

Update — we fitted the shade. Here’s what it was worth.

This piece ended on three recommendations, the loudest being shade the vessel, not the condenser. We took our own advice: an old beach shelter went up over the tub on 6 November. We split the metered history at that moment and ran the same analysis on both halves — same chiller, same tariff, same control code.

  • The afternoon peak dropped about 4 °C. The water used to rebound to 17–18 °C every day regardless of the overnight run; under shade it tops out around 14 °C. Heat gain was always the real problem, and this is the proof.
  • It now reaches setpoint overnight for the first time — 7.9 °C one morning, 7.0 °C the next, under the 8 °C target the plant had never met.
  • The afternoon “sun tax” is gone. Passive heat gain per degree of air-to-water difference used to run about 50% higher in the afternoon than the morning; post-shade that gap is effectively zero.

In the original article we modelled that halving the heat gain would let the overnight off-peak run alone reach 8 °C by 5:30am for about 13c. A shelter from the shed got us most of the way there, measured, in three days.

One number looks like a regression and isn’t. The chiller’s net afternoon cooling fell from about −0.44 °C/h to −0.07 °C/h — because the water now sits near 10 °C instead of 13.5 °C, close to the floor the plant can reach in warm afternoon air, so there’s simply less left to pull. A near break-even afternoon at 10 °C is a better outcome than a brisk −0.44 °C/h at 13.5 °C. Energy use rose slightly (about 3.6–4 kWh/day versus 2.5–3.3) for the same reason: the plant is now holding a genuinely cold bath instead of giving up by mid-afternoon. Most of that is free solar surplus plus the 8c overnight window, so it costs cents.

Two things we’re still watching. The shelter sits close to the condenser — exactly what we warned about — so we track compressor draw against air temperature; it’s holding around 310 W, unchanged, which means it isn’t yet re-breathing its own exhaust. And the forecast model still carries the pre-shade sun term, so it now predicts the 5:30am water warmer than it actually comes out; a refit against the shaded data is queued.

The lesson is the one this article pointed at from the start: fix the thermal fabric before you tune the controller. The decay curves and the forecast are what make the cheap overnight window land the water on target reliably — but the single biggest improvement on this system came from a cover that cost nothing, not a line of code.

Want this done at your place?

We build this sort of thing across Brisbane & South-East Queensland — properly, locally, and documented so you can keep it running.