A weather service predicts tomorrow's wind speed to the nearest knot. Your grid operations team receives that forecast and immediately knows it is not enough.
Wind speed does not translate linearly to megawatts. The relationship depends on turbine power curves that are nonlinear, array wakes that change with wind direction, and icing that degrades output. A sunny forecast does not mean the solar plant will hit rated capacity — cloud cover, dust on modules, grid curtailment, and sun angle all rewrite the story.
This gap costs money. Over-scheduling reserves means paying for unused capacity. Curtailing solar because you forecast shortfall but the sun comes out leaves revenue on the table. Under-scheduling reserves gets caught unhedged and forces expensive real-time balancing purchases.
Generation forecasts quantify confidence with a band around every point estimate, so dispatch and storage can be scheduled from day-ahead to intra-hourly horizons on what the grid will actually see. This is the renewable-generation forecasting capability in the RealAI energy stack: wind and solar output predicted across horizons so balancing, curtailment and storage decisions are made on what generation will actually do.
The Physics Between Forecast and Dispatch
A meteorological forecast is a distribution. A grid operation is a decision under uncertainty. The work of a generation forecast is to translate one into the other — and most of the difficulty lives in that translation.
Wind forecasts arrive as point estimates. The turbine's power curve is convex: output rises steeply from cut-in, climbs through rated power, then flattens as pitch control engages. A small error in wind prediction becomes a large error in generation on the steep part, nearly none on the plateau. A forecasting system that ignores this and reports symmetric bands is quietly lying about its own uncertainty.
Array wake effects deepen the nonlinearity. Wind hitting one row of turbines exits degraded for the next. Available wind speed at each turbine depends on wind direction and array geometry. Two farms with identical turbines and forecast wind can produce materially different power simply because their layouts shadow wind differently.
Solar carries parallel physics. A cloud-free irradiance forecast leaves open module temperature, which climbs under full sun and drags down efficiency. Dust, soiling, and panel angle all pull output down. Curtailment and reactive-power obligations mean the plant's actual electrical output is often less than physics alone allows. A model that learns from raw output data without accounting for curtailment will mistake an operator's instruction for a weather signal.
The forecasting problem is generation, not weather. The shape is Bayesian: "given all the physics, the historical outputs at this wind speed and direction, and current conditions, what is our credible range of tomorrow's generation?" That framing — a confidence interval attached to every forecast — is what makes the output usable in a control room, where the question is never "what is the number" but "how much should I hedge against the number being wrong?"
Horizon-Aware Confidence Bands in Production
Production systems separate forecasts by horizon because the drivers of error change as you approach real time.
Day-ahead. Forecast error here is dominated by meteorological uncertainty. Large-scale weather patterns shift and fronts move slower or faster than predicted. Confidence bands are widest. Reserve schedulers decide whether they hedge toward a higher percentile — because they know exactly what underestimating wind costs when they buy it back in the real-time market.
Intra-day, a few hours out. Weather-model error shrinks, but ramp uncertainty emerges. Wind can gust sharply as a storm's outflow reaches the farm. Solar can fall away in minutes if clouds cross the site. The confidence band changes shape — dispatch now cares about "how fast could it change" rather than "how much on average." Expensive surprises live in the tails.
Intra-hourly nowcasting. Near real time, meteorological error becomes noise compared to the ramp itself. You can nearly see the cloud and track the gust front on radar. The forecast becomes extrapolation — if the current ramp is steep and no new weather appears on radar, what are the odds it continues? Bands narrow sharply and often become asymmetric.
These systems run separate models per horizon: one calibrated against day-ahead error, one against intra-day ramps, one against nowcasts. The error characteristics are too different to share one set of parameters.
Physics-Informed Models: Holding the Edge Across New Farms
A standard statistical forecast trained on a single wind farm's history performs brilliantly for that farm. Swap in a different site, and accuracy craters.
The reason is structural. The relationship between weather input and electrical output is unique to that farm's geography, turbine model and terrain. A black-box regression learns all of this implicitly and cannot tell which parts transfer to a new site.
Systems that hold accuracy across new sites encode the physics that does not change. Instead of one opaque regression, they are built in layers. A meteorological input layer carries weather-model output with explicit terms for wind components relevant to the site's terrain and solar angles accounting for elevation. A power-curve layer applies a parametric model of the turbine's published curve, with site-specific icing, soiling and temperature losses fitted as correction factors — this layer does not learn the power curve, it constrains the model to obey it. An uncertainty layer wraps everything in calibrated Bayesian bounds.
This layered structure is what makes the RealAI energy work physics-informed: predictions are constrained by physical priors rather than free to fit any pattern in the data, earning the trust of teams who must bet a budget on the output. When you move such a model to a new farm, the meteorological input changes with new terrain. The power curve stays — same turbine manufacturer, same physics. Only site-specific icing and soiling corrections need to refit on new data. The model's accuracy at the new site can approach the original's not because it re-learned everything from scratch, but because the physics-informed structure transferred.
For an operator managing a portfolio, that transfer is the whole point. A new hybrid site comes online; instead of waiting a full season for local-data accuracy, the system ships day-one with physics-based predictions that are immediately actionable, then improves automatically as the farm's own data accumulates.
- ~20%
- More value from wind via day-ahead forecasting (DeepMind, 2019)
- ~13%
- Lower solar forecast error with ML (illustrative)
- Calibrated
- Bayesian bounds on every forecast
- 4–6 weeks
- To a ranked roadmap
Closing the Loop: Recalibration Against Reality
Generation forecasts decay slowly and quietly because the model keeps producing confident-looking numbers long after the conditions it was trained on have moved.
Seasonal weather patterns shift. Turbines degrade through blade erosion and bearing wear. New renewable capacity reshapes curtailment patterns. Load reshapes as electrification advances. Every one of these moves the relationship between weather and delivered megawatts.
Systems that stay accurate run a scheduled recalibration loop. On a short cadence, they compare actual generation against the forecast for every site and time window, computing bias and dispersion. On a longer cadence, site-specific correction factors are retrained. At each season change, ramp patterns are revalidated. Whenever the world changes structurally, the forecast is validated under new conditions before it is trusted.
The hardest part is honest calibration. A model is only well-calibrated if, over many observations, that stated share of actuals really does fall inside the band. If actuals consistently land above the upper bound, the model is over-cautious and dispatch wastes reserves. If actuals frequently break outside, the model is over-confident and dispatch gets caught unhedged. Calibration is the difference between a band a scheduler can size a decision against and a band that quietly lies.
The recalibration process scores calibration explicitly — does the median forecast actually sit at the median of outcomes, is the stated tail risk the real tail risk — and corrects it. Not by re-tuning physics, but by adjusting ensemble spread and site-specific loss factors. This closed loop keeps the model honest as the grid changes.
A weather forecast is not a generation forecast. The confidence band between them is where operations happen.
Why It Shipped: Uncertainty as Operational Necessity
Operators stop asking for point estimates and start asking for confidence. In capital-intensive, safety-critical operations, a calibrated "how sure are we" is what makes a model deployable.
A reserve scheduler needs the median and the band around it, because the band tells them how much ramping reserve to schedule. If they schedule for the median, they are under-covered roughly half the time. If they schedule for the high end, they are over-provisioned much of the time. The band lets them choose their risk tolerance deliberately.
A storage operator needs not only when solar will peak, but the ramp shape — how steeply generation climbs — because that tells them when to start charging the battery. A forecast that reports only the peak hides exactly the information a storage dispatch decision needs.
The design choice that makes these models deployable is that they are built around uncertainty from the start — not accuracy first, with uncertainty bolted on as an afterthought. Every part exists to quantify how sure the system is. That is what earns enough trust from dispatch teams to actually change how they run the grid.
Where to start
The assessment phase runs in four to six weeks to a ranked roadmap.
Start by inventorying what you have: historical weather data — is it high-resolution and sub-hourly — generation telemetry that may or may not capture curtailment and reactive power, and the operational decisions that depend on it, from reserve schedules and dispatch logs to storage patterns and real-time imbalance prices.
Map your top operational pain points. Is it the cost of over-scheduling reserves? Curtailment because you under-forecast, or missed revenue because you over-forecast? Real-time balancing cost when intra-hourly ramps come in worse than expected? Score each by capital at risk and data quality. The highest-impact, data-ready use case becomes the pilot.
The Transform phase builds a physics-informed generation model on that best-data use case, piloted against the horizons with the most operational leverage. The model ships with uncertainty bands baked in, and dispatch teams begin using the bands rather than the point estimate. Once operational data accumulates, the model is validated: are the confidence bands holding? The Sustain phase then locks in the recurring bias check and recalibration, so that when the model drifts — as weather regimes shift, gear wears and the grid changes — you catch it with data rather than intuition.
The signal flow is from physics to operations to ground truth, and back to retrain. That loop, more than any single accuracy figure, is what keeps supply real.
“A weather forecast is not a generation forecast. The confidence band between them is where operations happen.”
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.