Steel Mills as Batteries: A Clever Shortcut That Lets Factories Balance the Grid

Here's the thing about the grid of the future: it's going to be full of contradictions. Solar and wind power, which now supply a growing share of the world's electricity, are intermittent by nature — they peak when the sun shines and the wind blows, regardless of when we actually want power. The most flexible way to smooth out those swings isn't a giant battery. It's industry.

Steel mills, cement plants, and powder factories are enormous electricity consumers that can, in principle, shift their energy use to whatever hour the grid needs it most. But there's a catch buried deep inside every industrial facility: the physics of production. You can't just tell a steel plant to use more power at 3 a.m. — the constraints of furnaces, grinders, and process schedules are tangled, nonlinear, and studded with on/off decisions that make their energy-use models nightmarishly complex. These "mixed-integer" models are so hard that grid operators, who run electricity markets every single day, essentially can't use them to decide how much power to buy.

A team from Tsinghua University and Imperial College London has figured out a way around this wall. Rather than trying to compress those gnarly industrial constraints with pure math, they let the data do the talking. Their approach — called D3R, for data-driven dimension reduction — watches how an industrial facility actually uses energy in response to changing prices, then reverse-engineers a simple, grid-friendly model that mimics that behavior. In tests across three real industrial plants — a cement plant, a steel powder plant, and a steelmaking facility — the new method cut the modeling error roughly in half or better compared with the best available analytical techniques, hitting error rates as low as 3.6% (Lyu et al., 2026). That's the difference between a model a grid operator can actually use and one that's hopelessly complicated.

The Science

To understand what D3R does, you first have to appreciate what's standing in the way. Industrial energy models — built with techniques called Resource-Task Networks (RTNs) or State-Task Networks (STNs) — represent every step of a production process: which machines run, for how long, in what order, subject to what physical limits (Lyu et al., 2026). These models need "integer" variables to capture binary decisions: a furnace is either on or off; a production line either runs or it doesn't. Mixed-integer models of this kind are notoriously hard to solve, and impossible to fold directly into an electricity market clearing process that must reach a decision for thousands of loads in seconds.

The traditional workaround is dimension reduction: replace the complex, high-dimensional constraints with a simpler "lower-dimensional" approximation — a box, an ellipse, or a geometric shape called a zonotope — that roughly captures what the original model would allow. These are called analytical methods, and they have two fundamental weaknesses. First, the pre-chosen geometric template rarely fits every kind of industrial load; a cement plant and a steel mill have utterly different constraint structures. Second, these methods require the original constraints to be convex — mathematically smooth and well-behaved. Mixed-integer constraints, by definition, are not (Lyu et al., 2026).

The D3R insight is beautifully simple: instead of guessing a geometry and deriving parameters from assumptions, train the reduced model the way you'd train a machine-learning model — on data. The researchers took the optimal energy-use records of industrial loads (the actual hourly consumption patterns that minimize cost under real electricity prices), and then asked: what simple model, when given those same prices, would reproduce those same consumption decisions? That's a problem called inverse optimization — instead of "given constraints, find the best behavior," it's "given observed behavior, find the constraints." The output is a fitted, low-dimensional linear model that captures the facility's flexibility without needing to model its furnaces.

As the framework illustration shows, the D3R pipeline takes the original high-dimensional constraints (represented by the industrial process model), feeds them optimal energy-usage data, and outputs a low-dimensional constraint set that approximates the original behavior (Lyu et al., 2026)

.

The team chose a specific low-dimensional form to train: the adjustable load fleet (ALF) model. It treats a factory as a small number (1 or 2, in their tests) of "adjustable loads," each with a power rating (a floor and ceiling on how fast it can consume) and an energy budget (a floor and ceiling on total daily consumption). That's it. Three physical truths — how much equipment there is, how fast it can run, and what the daily production target requires — rendered as linear constraints (Lyu et al., 2026). No furnace physics. No integer decisions. Just a handful of numbers that capture the same flexibility.

What They Found

The results are stark when laid side by side against the status quo. The researchers tested the method on three datasets — a cement plant, a steel powder plant, and a steelmaking facility — using real PJM electricity market prices from July 2022, training on the first 21 days and testing on the final ten. They compared D3R against the two main existing approaches: the simple AL model (SAL), which just reads off power limits from historical data, and the optimal virtual battery model (OVB), the leading analytical inner-approximation method.

The headline numbers speak clearly. On the cement plant, D3R achieved an error of 8.9% versus 10.2% for SAL and 25.4% for OVB. On the steel powder plant, the gap widened dramatically: D3R got down to 10.3% while SAL stalled at 23.0% and the analytical OVB method limped in at 36.7%. And on the steelmaking plant — the hardest case, with over 10,000 integer variables in the full model — D3R achieved just 3.6% error, while the analytical method couldn't even be applied at all (Lyu et al., 2026)

Dimension-reduction error on the cement plant

Label	Value
Cement (nRMSE)	25.4
Cement (SAL)	10.2
Cement (D3R-1)	8.9
Cement (D3R-2)	9.5

.

For context on those numbers: all of this happens in a world where grid flexibility contracts routinely tolerate a 10% to 20% deviation from scheduled consumption. D3R's reduced models land comfortably inside that band — in fact, the steelmaking model at 3.6% blows right past it.

The method doesn't just win on accuracy; it wins on survival. The OVB analytical method failed outright in two of the three test cases — it couldn't handle the integer-heavy steelmaking model at all, and for the steel powder plant (a 10-stage production process) it failed to converge even after 24 hours of computation, forcing the authors to settle for whatever accuracy it had reached when they gave up (Lyu et al., 2026)

Dimension-reduction error on the steel powder plant

Label	Value
Steel powder (OVB)	36.7
Steel powder (SAL)	23
Steel powder (D3R-1)	17.2
Steel powder (D3R-2)	10.3

. D3R, by contrast, converged on every dataset.

One of the more striking findings is what D3R's convergence looks like. The training curves show the fitted parameters settling smoothly toward their final values, and the method's computation time stays flat even as you feed it more days of data — the algorithm randomly samples a small "batch" of days at each step rather than chewing through the entire history (Lyu et al., 2026)

.

Why This Changes Things

Stick with the practical problem for a moment, because the stakes are bigger than a benchmark table. Electricity grids worldwide are being asked to do something they've never done: integrate vast, unpredictable renewable energy at scale while keeping the lights on. The cheapest, fastest solution available isn't more batteries — it's using existing industrial flexibility better.

But industry can only participate in markets like demand response, reserves, and frequency regulation if its flexibility can be represented in a form grid operators can compute with, on the spot, alongside thousands of other participants. Today, that's often impossible. An industrial demand-response aggregator might control a portfolio of factories, each with a mathematically intractable model, and have no clean way to tell the market "here's what this fleet can do tonight." The analytical dimension-reduction tools that exist fail precisely on the hardest, most valuable loads — the ones with the most complex production processes — and the ones with integer constraints (Lyu et al., 2026).

This is where D3R inverts the entire paradigm. The paper's reference to an "industry tolerance" of 10% to 20% is the key philosophical unlock: strict mathematical equivalence between the original and reduced models is not only computationally infeasible, it's unnecessary. Markets don't need perfection; they need a model that's good enough, fast enough, and honest about what the flexible resource can actually do.

There's also a deeper point about how we build models of real systems. The conventional approach is bottom-up: start from physics, derive equations, then simplify them analytically under assumptions that may not hold. D3R is top-down: start from observed behavior, then find the simplest model that explains it. This isn't novel in machine learning — but it's essentially unheard of in the world of industrial load modeling, which has been dominated by analytical derivation for decades. The authors make the point explicitly: many of the analytical methods fail not because the underlying plant is too complex, but because their pre-specified geometric templates "lack adaptability across various industrial loads" (Lyu et al., 2026).

That said, honesty demands noting the costs. Choosing the ALF form does throw away some real physics — the model "may overlook some temporal coupling between production processes and relax integer variables" (Lyu et al., 2026). For most market applications, that's an acceptable trade; for a plant where precise process timing is safety-critical, it might not be. D3R also assumes you have data that reflects genuinely cost-optimal decisions. If a factory uses a sloppy heuristic schedule instead of optimizing, the observed data won't cleanly reveal its true flexibility — although the authors note the framework can handle this by first simulating optimal data under varying prices before training.

What's Next

The most exciting part of D3R might be how general it is. The paper is about industrial loads, but the framework is really a way to compress any complex constraint set using only input-output data. The authors are explicit that the ALF model is just one choice of reduced form — "the D3R framework is not limited to the ALF model" (Lyu et al., 2026). As they see it, future work should chase more realistic reduced-constraint forms that capture industry behavior even more faithfully, and figure out how to decompose a fleet-level energy schedule back down to individual machines — so you don't just know the plant should consume 40 MWh tonight, but which grinder should handle it.

There are real questions left open. The three test plants are real, but the data feeding them were simulated — the "observed" optimal behaviors came from running the original models, not from meters in a factory (Lyu et al., 2026). Wiring D3R up to live industrial data, with all its noise, missing values, and operator quirks, is the natural next stress test. And there's the question of how a fleet of D3R-reduced models behaves when they're all responding to the same price signal at once — the coordination dynamics that markets live and die by.

But the direction of travel is clear and genuinely hopeful. As renewable penetration grows, the grid's ability to absorb fluctuation increasingly depends on the demand side flexing as much as the supply side. For that to happen at industrial scale, we need models of industrial flexibility that are both accurate enough to trust and simple enough to compute — and for the first time, we have a method that delivers both, on the exact workloads where everything else falls over. When the steel plant's flexibility can finally be packaged in a form a market can price, industry stops being a passive drag on the grid and starts being one of its most powerful balancing tools.

That's not just an engineering convenience. It's one more piece of the puzzle for a decarbonized grid that works.