DeepSYS cooling station energy savings: from logging records to the first report

Last month I helped a friend doing a park renovation look at air-conditioning electricity bills. He spread the past half year's bills on the table and said the electricity cost kept rising, and the property management couldn't say which equipment was eating the power either. Chillers, pumps, cooling towers — he knew all the names, but who was wasting and how much, nobody could answer. That's also why I later went to try DeepSYS — what it manages is how the whole building's air-conditioning system runs together most economically.

This thing is worth using, but after installation it won't take effect on its own; you have to lay out your assets first. If the ledger isn't fully entered, no matter how accurate the model is, it's useless. Whether the installation threshold is high mainly hinges on this ledger-entry step.

In plain terms, a building's air conditioning is several pieces of equipment restraining each other. The chiller produces cooling, the pump sends chilled water to each terminal, and the cooling tower discharges the heat. Tuning each one individually to its highest efficiency doesn't necessarily add up to the most economical overall — it might even use more power. This is completely different from one AC and one remote at home.

DeepSYS's approach is to first replicate this system in software. It uses mechanism plus data-driven modeling, building models at the equipment level, and the materials say prediction error is controlled to RMSE under 3%. You can understand RMSE as the average difference between what the model guesses and the real value; within 3% is good enough for a slow variable like air conditioning.

Then it tries combinations in this virtual model. It's said one optimization run has to test more than 100,000 combinations, find the solution approaching the energy-saving limit, then send it down to the real equipment, fully automatic, no one watching.

The renovation case at Sichuan Guangda Pharmaceutical had a 23.6% energy-saving rate. That's someone else's project, not a number you get just by plugging in — keep the right mindset.

By the way, a word on the industry size. Industrial and building energy consumption accounts for 70%, of which general equipment like HVAC and air compressors accounts for 30% to 50%. The ceiling is indeed high, so there have been quite a few teams working this area in recent years.

I went through it myself from zero. First open an account and get project permissions — this step is usually opened for you by the integrator or vendor; you can't register yourself, because it hangs under a project. Once you have it, go into the backend and check whether there's a site list on the left. Then create a site: click new site, fill in the building name, city, building area, and air-conditioning system type. Don't fill in the city carelessly — it's tied to the later weather data; fill in the wrong city and the load model will be off as a whole.

Next is entering the equipment ledger, the most time-consuming step. Enter chillers, pumps, and cooling towers one by one, filling in model, rated power, rated cooling capacity, and whether it's variable-frequency. If there's a nameplate, copy from the nameplate; don't estimate from memory. Click add equipment, select equipment type, save after filling, and an equipment card will be generated on the interface. Then attach metering points, mapping the measuring points of electricity meters, water temperature sensors, and flow meters one-to-one to specific equipment. When a real-time curve appears under each equipment card, the data is through. If the curve is flat or stuck at zero, the points are wired wrong — go back and check. Finally run modeling and optimization: click model training, wait for it to finish, and it will produce a baseline report telling you how much this system could theoretically save. Note it's theoretical, not a promise.

On my side the first round took a bit over two days, and a day and a half of that was entering the ledger. When I saw the report after running it, honestly I wasn't very excited, because the number it gave was more conservative than I expected.

There are three pitfalls that are easy to step into. First, cutting corners on the ledger and using rough numbers — if rated power is off by one tier, the optimized strategy will be off, and it won't report an error, it'll just quietly give you a not-good-enough solution. Second, metering points mismatched with equipment — if the pump's electricity meter is connected to the cooling tower, the curve still shows numbers, the model learns the wrong thing, and you won't easily notice; after entering, you must check the correspondence between points and equipment item by item. Third, wanting to turn on full automation right away — run suggestion mode first, see how far the strategy it gives differs from the on-site veteran's judgment, and only go closed-loop once they match.

For the next step I suggest bringing the air compressor system in too. Air conditioning and air compression are two independent systems, but in many parks the electricity bills for both go through the same meter. Optimizing separately makes it easy to save on one side and increase on the other; only looking at them together is the real account.

Going forward, this kind of system will probably shift from an energy-saving tool to an operations entry point. Saving power is only the first layer; its bigger use is pulling equipment ledgers, real-time data, and operating strategies together. Later, replacing equipment, doing carbon accounting, even dispatching repair orders, can all grow out of this data. Whoever gets the ledger entered cleanly first will save themselves more trouble later.