Your grid is shaped like a duck, and the shape is getting more exaggerated every year.
At midday, utility-scale solar floods the grid with cheap, abundant power, and demand for conventional generation plummets because the sun is doing the work. Then sunset hits. Solar generation drops away. Evening demand climbs as people come home. Your conventional generation, your storage, your demand-response programs all have to ramp hard and fast to fill the gap. The overnight hours settle into a valley. Plot net load over a day and the shape resembles a duck: a low belly at midday, a steep climb on the evening ramp, a head rising toward peak.
That duck curve is no longer a curiosity for grid planners. It is the defining operating challenge of a grid where renewable penetration keeps climbing. And the grid operators managing it are doing it in real time, with decisions that compound minutes apart.
The operators winning are not guessing the curve. They are reading it.
The Duck Curve Is a Dispatch Problem Wearing a Generation Problem's Mask
Grid operators have managed peaks for decades. A hot summer afternoon. A winter cold snap. You forecast demand, size your generation to cover it, and the economics are predictable because demand is relatively smooth and slowly moving.
Renewables broke that assumption. A cloud bank rolling across a solar farm can drop generation quickly. Wind output can swing within an hour. These swings are not driven by demand; they are driven by weather and time of day. And because solar and wind are zero-marginal-cost resources, they displace expensive conventional generation whenever they are available. The moment the sun rises on a clear day, the marginal price falls. Utilities curtail their expensive plant to avoid taking a loss on every megawatt-hour. Then sunset comes and the curve snaps upward.
The challenge for a grid operator is not predicting demand anymore. It is predicting the shape of net load, which is demand minus renewable generation, hours into a future that is being written by cloud cover, wind shear and the sunset.
If you read it wrong, the costs are unforgiving. Commit a fast-ramping peaker unit you did not need and you bleed money on no-load costs. Under-commit and you are buying expensive ancillary services to keep frequency and voltage stable as you scramble to ramp. Mistime your battery charge and you hit peak load with an empty tank and a thin reserve margin. The operators who hold margin and manage cost are the ones who know, hours ahead, what the duck curve actually looks like.
The Physics-Informed Read: Renewable Forecast + Demand Pattern + Grid Constraints
The duck curve is not random. It is the sum of three signals.
Renewable generation is driven by solar irradiance, wind shear and temperature. A physics-informed model learns the relationship between recent satellite-derived cloud cover, numerical weather forecasts and the solar output a utility actually sees at each farm. The same holds for wind, where blade pitch, rotor speed and anemometer data fold into a weather-informed forecast. These models do not guess; they reason from the physics of how sun and wind interact with the grid hardware. That physics grounding is what earned the trust of the people who have to act on the output — the same way a 3D convolutional architecture constrained by geological priors and thermodynamic loss functions earned the trust of exploration teams betting a drilling budget.
Demand follows patterns. The morning ramp as people wake and plug in. The daytime plateau punctuated by midday swings. The evening spike as people come home. The overnight trough. These patterns are learnable from years of interval meter data, but they shift with temperature, with day of the week, with holidays and grid events. A model that reads recent demand history alongside a weather forecast and a calendar is far more accurate than the seasonal averages utilities have leaned on.
Grid constraints are the transmission limits, losses and physical laws that bound how fast you can move power across the network. A properly constrained model respects those bounds; an unconstrained one will predict a beautiful net-load shape that is physically impossible to deliver.
Put those three together — renewable-generation physics, demand patterns and grid constraints — and a model reads the duck curve not as a surprise but as a consequence of knowable forces. It tells you when the solar peak arrives, how steep the evening ramp will be and how much reserve you need to hold stability margin through the rise. That is a decision you can act on.
Bayesian Uncertainty: The Forecast That Tells You How Sure to Be
The forecasts grid operators actually use in production are not point estimates. They are distributions. A model does not just say the evening ramp will be a given size; it attaches a confidence interval to that number. That band is the instrument. It tells you how much risk you are taking if you commit less ramping capacity than the upper bound. It tells you where over-commitment starts bleeding you on no-load costs. It tells you where the real operating point sits — inside the band, but close enough to make a bet.
Without that confidence interval, a forecast is a guess with a number attached. With it, it is a risk-aware decision tool. This is the same design choice that made the energy practice's models deployable in the first place: a Bayesian layer attaches a confidence interval to every reservoir and load forecast, and in capital-intensive, safety-critical operations a calibrated "how sure are we" is what makes a model usable, not raw accuracy alone.
The models that earn trust in dispatch centers share three traits:
-
They update frequently — digesting the latest cloud satellite, wind data and demand reads rather than running stale on a forecast made hours ago.
-
They degrade gracefully — when uncertainty bounds widen because conditions are volatile, the model surfaces that rather than projecting false confidence. Operators respect a model that says it is less sure today more than one that pretends certainty it does not have.
-
They explain their errors — so that when the model misses, the operator can see whether a cloud-cover error or an unexpected demand shift drove it, and decide when to trust the next forecast and when to override it.
The DuckCurve Instrument: Wired Into Dispatch
The highest-value implementations tie the forecast directly into the operating plan. The DuckCurve instrument reads net-load shape in real time and sits upstream of the unit-commitment decision that decides which generation to turn on, when to charge or discharge storage, and which demand-response programs to activate. The instrument publishes its forecast and its confidence intervals; the unit-commitment logic reads those intervals and commits capacity tall enough to cover the high end of the ramp without burning money standing idle. The battery runs a charge schedule that prepares for the peak, informed by the model's read of when the peak will hit. Demand-response dispatchers see the forecast and pre-stage programs, knowing a hard ramp is coming and that response can blunt it.
Operators do not live in the model. They live in their normal dispatch console. But the console now lights up with a picture of what is coming, and dispatch acts against it instead of chasing it. That is the shift from real-time firefighting to staying ahead of the swing.
The grid operator who reads tomorrow's duck curve today does not chase the swing — they lead it.
- Real-time
- Net-load forecasting
- Horizon-aware
- Output forecasts
- 4–6 weeks
- To a ranked roadmap
Data Readiness: You Probably Have More Than You Think
The trap that stalls most grid operators is assuming they need perfect data to start. In reality the data is already there, scattered across SCADA systems, smart-meter backends and renewable-farm archives.
- SCADA dispatch data reports power flow across transmission lines, generators and load points. Years of history give you the ground truth of net load over conditions you have already seen.
- Renewable generation telemetry is the training target: utility-scale solar and wind farms export meter data to the grid operator, so you know what generation looked like on the sunny days and windy nights already behind you.
- Demand comes from smart meters at the feeder level, and that data combined with SCADA readings at aggregation points tells you the shape of demand at sub-hourly resolution — which is enough; you do not need real-time household data.
- Weather comes from publicly available numerical weather predictions plus satellite-derived cloud cover. Tie these to your renewable output and you have the feature set a physics-informed model needs.
The energy practice treats this heterogeneous reality as the starting point rather than a blocker. Decades of data collected with different methods, sparse logs, and SCADA and smart-meter histories are exactly what the assessment phase is built to inventory. The phase takes stock of what you have, assesses data quality and identifies the gaps, and the model is built on what is reliable. A common finding is that solar history is rich while wind is spotty, so the roadmap starts with the renewables you can forecast well and adds the rest as the archive grows.
Where to Start
The assessment is a 4–6 week audit of your grid's duck curve, your renewable footprint and the forecasting edge you can unlock. It inventories the heterogeneous reality of energy data and ranks where physics-informed AI changes a real decision, producing a roadmap of exploration and grid use cases scored by capital at risk and confidence achievable.
You start by understanding your duck: mapping your utility's actual net-load shape across seasons, finding the sharpest ramps and the hours where forecast error costs the most. You audit your data sources — SCADA history, renewable-generation telemetry, demand archives and weather feeds — and quantify their quality. You map which generator types are the bottleneck, whether your peakers are the constraint or your ramp rate is, and whether your storage charges at the wrong time because the forecast is absent.
The output is a short, ranked roadmap of improvements, typically starting with renewable-generation forecasting because the data is cleanest, then demand patterns, then an integrated net-load model that accounts for grid physics. Each item is tied to your utility's actual economics: cost reduced on balancing services, margin protected on the evening ramp, certainty gained in unit commitment. From there the build phase hardens the net-load model into production, where operators act on confidence intervals rather than point estimates and the DuckCurve instrument ships wired into dispatch. Then sustain closes the loop, recalibrating duck-curve forecasts as the generation mix shifts and monitoring uncertainty bands so the model stays honest across a changing grid.
And for the first time, operators stop chasing the curve, and start leading it.
“The grid operator who reads tomorrow's duck curve today does not chase the swing — they lead it.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production — sovereign models, governed data, and agents you can audit. Start with a value-first assessment.