By Chetan Chaudhari | Call me old-fashioned, but I still use my six-year-old phone. The same phone that used to get me through a commute, a full day at work, several hours of Spotify, and sometimes even two days on a single charge. Now it’s a miracle if it survives a day out without a battery bank in my pocket, and I’m left hoping I don’t get stranded with no maps, no email and (gasp) no Spotify.
Nothing about the phone is broken. The battery degrades every year, iOS updates keep increasing computing load (I am looking at you Liquid glass!), and my days keep getting busier — despite the AI making things faster. The “all-day battery” spec was true. It was true for a new phone on an average day.
Hybrid power parks being built for AI data centers come with a similar promise: Solar and storage will cover the load around the clock, and gas will fill in only now and then. And, hopefully that interconnection will come through soon and all will be well.
In year one, that promise mostly holds. The question that matters is what year ten or twenty looks like, when the panels have aged, the weather has done what weather does – which is out of our control, and the battery is cycled more often than anyone planned.
Hybrid power is speed to power
AI data centers need a lot of power, and they need it now. Waiting years for a grid connection isn’t an option when the servers are already on order. More and more, the answer is to build the power park next door: a large solar array, a battery big enough to carry the load into the night, and a gas turbine as backup for the hours the other two can’t cover.
It’s an elegant setup. The sun provides daytime capacity and charges the battery, the battery covers the evening, and the gas turbine comes on only when it has to. Contracts get written around it. Sustainability pledges are made around it. Financing gets sized around it.
Every day isn’t Groundhog Day
While designing our elegant solution, estimating a solar plant’s output depends on more than just the weather. It also depends on how good the physics models (PVsysts of the world) that estimate conversion of sunlight into electricity are. And it depends on dozens of inputs we can only estimate: how fast the panels will degrade, will they respond to temperature the same across the board, when the battery will need fresh capacity added. Each input is really a reasonable guess with a range around it.
Let every one of those inputs vary across its realistic range, and a single day of solar output stops being one definitive estimate. It becomes a whole bundle of possible paths (Fig. 1a). Some days are close to what one would call a typical-year forecast. Some aren’t close at all due to interacting uncertainties.
In a hybrid plant, that spread amplifies further. Solar is the charger for the battery. A weaker day means less energy stored, which means an emptier battery at sunset, which means the gas turbine comes on earlier. Fig. 1b shows the same bundle of possibilities passed through to the battery: thousands of ways it might charge and drain over the same two days.
Making design decisions based on a typical generation year gives you just one of those lines (a typical possibility) and limits the robustness of the design.
Fig. 1. (a) Possible solar output over two days with weather and modeling uncertainties are accounted for. (b) The resulting range of state of charge levels or SOC. Each gray line is one plausible outcome.
A 100 MW data center, 20 years of weather
To see how much this matters, let’s consider a hybrid plant serving a 100 MW data center. The plant pairs 300 MW of solar with a 300 MW, four-hour battery, plus a 100 MW gas turbine as backup. Every hour, the plant is dispatched to keep gas use as low as possible.
Then we ran it two ways. The conventional way uses one typical weather year and standard degradation assumptions. The uncertainty-aware way lets weather and dozens of modeling inputs vary across their realistic ranges for every year of operation.
In year one, the two approaches roughly agree. The conventional model says the data center runs on clean power about 83% of the time. The uncertainty-aware P50 lands in the same neighborhood, at 80–84%. A bad-weather year costs about two points. So far, so good: this is the new phone.
By year twenty, the two have split apart (Fig. 2). The conventional line has barely moved, still at about 81%. The uncertainty-aware P50 has slid to about 69%, with a likely range of 58% to 80%. That 12-point gap equals roughly 1,000 extra hours a year when the data center runs on gas instead of clean power.
And this analysis is the generous version of it. We treated the battery, turbine and load as perfect, with no losses, heck the battery isn’t even degrading, so the effect of solar uncertainty would show on its own. Real equipment losses would push those numbers lower, not higher.
Fig. 2. Share of hours the data center runs on clean power, by operating year. The conventional typical-year forecast (purple) stays nearly flat. The uncertainty-aware P50 and P90 (blue and red) decline as uncertainties compound, and the shaded band shows the likely range.
How to make promises that can be kept
None of this is an argument against solar or for that fact, hybrid plants. They’re still the fastest way to get large amounts of clean power capacity to the data centers that need it. It’s an argument for knowing what you’re promising before you sign.
For owner-operators, the clean-energy share which in many ways is a proxy for effective “clean” capacity factor is no longer a marketing number. It shows up in contracts, sometimes with penalties attached. It shows up in the community meetings where you are assuring the communities of the minimal environmental footprint. A financial agreement priced on the conventional line gets harder to meet every year. Long-term owners, not the forecast, ends up covering the difference, owning the multi-fold risk.
For lenders and offtakers, P50 and P90 over the plant’s whole life show what’s actually bankable. An investment that can pay for itself. A year-one number, or a typical-generation-year number degrading at a nominal rate for twenty years, doesn’t.
For engineers, the decisions themselves change with the dimension of uncertainty in mind. How big should the battery be? When should capacity be added? How much solar overbuild is worth it? How much gas backup is enough? Sized for the average, the optimistic answers look one way. Sized for the realistic spread, the answers will be robust.
The uncertainty is always there. The only question is whether it gets priced before the start of the project or paid for later, often compounded.
Year six, year 20
I’ll probably give in and replace my phone soon. That’s the nice thing about a phone: when the battery can’t keep up, you buy a new one.
A long-term asset such as a hybrid power park doesn’t get that option. The plant is built once and the contract is signed once, and both have to hold up in year twenty or succumb to unfavorable terms during hand-offs. So before the ink dries, run the plant through the numerous possibilities, not just one typical one. The typical year is a fine place to start. It’s a risky place to stop.
Author bio: Chetan Chaudhari, is the CEO of PowerUQ an uncertainty quantification platform built for solar energy yield assessment and a principal at Sierra Analytica. He brings over 18 years of experience building data science, software and hardware tech for solar, storage and smart grid projects.
— Solar Builder magazine
Source link