When a solar developer submits a production forecast to the market, P50 is the number that matters to them — expected energy, expected revenue, the basis for their hedge. Grid operators live in a different decision regime. For a balancing authority setting spinning reserve the night before a high-solar afternoon, the relevant question is not "what will the solar fleet probably generate?" It is "what is the worst credible outcome I need to be prepared to cover?"
Those are structurally different questions, and they call for structurally different forecast outputs. Using a P50 forecast to size reserve is a category error that gets masked by the average but surfaces as a frequency-response event on the tail days.
What P50 and P90 actually mean in a forecast context
In energy forecasting convention, a P50 value represents the median expected production — 50% of ensemble realizations fall above it, 50% below. A P90 value means 90% of ensemble realizations are expected to be at or above that level; only 10% of outcomes are lower. Note the directionality: in the solar production convention, a higher P number means a more conservative (lower production) estimate, because the tail risk for grid reserve purposes is under-production, not over-production.
This is worth stating plainly because different industries flip the convention. In financial risk, P90 revenue is an optimistic projection. In solar production forecasting for grid operations, P90 output is the pessimistic tail — the number you plan against when you need to guarantee adequacy. Mixing these up in a reserve commitment workflow has real consequences.
The dispatch economics of the wrong choice
Consider a balancing area with 800 MW of utility-scale solar dispatched in the day-ahead market. At noon the following day, the P50 forecast is 680 MW and the P90 forecast is 580 MW — a 100 MW spread driven primarily by cumulus convective cloud uncertainty over a high-altitude plateau site southeast of the urban load center.
If the operator commits reserves based on P50, they carry enough spinning reserve to cover a deviation from 680 MW. If the actual output tracks the P90 scenario — cloud cover arrives two hours earlier than the median model run predicted — the operator is short 100 MW of fast-response capacity. That shortfall has to be covered by non-spinning reserve or emergency interchange, both of which are more expensive and slower to ramp.
The inverse is equally real. If an operator habitually commits reserves against the P90 tail on every interval, they are systematically over-procuring spinning reserve by the spread between P50 and P90 — roughly 12–18% of committed thermal capacity on days with moderate forecast uncertainty. That over-procurement is real dollars: a mid-size balancing authority paying a combined-cycle unit $8–12/MW-hr to spin idle carries a meaningful cost per over-committed GWh.
We're not saying P90 commitment is always correct and P50 is always wrong. We're saying the choice between them should be a deliberate function of uncertainty spread, ramp exposure, and reserve margin headroom — not a default to whichever number the forecast vendor sends first.
When the uncertainty spread is the real signal
The actionable insight from a probabilistic forecast is not the P50 or P90 value in isolation — it is the spread between them. On a cloudless high-pressure day, a 400 MW solar installation might show a P50-P90 spread of 15 MW across the afternoon: low uncertainty, model agreement is strong, commit against P50 with minimal reserve add. On a frontal passage day, the same asset might show a P50-P90 spread of 180 MW at the 3-hour-ahead horizon: high uncertainty, the ensemble members are widely divergent on cloud timing, carry more reserve.
This is the argument for probabilistic forecast bands over deterministic single-point outputs. A single-point forecast cannot communicate that distinction. It presents the same confidence implicitly regardless of whether the NWP ensemble is tightly clustered or widely spread. The dispatch engineer has no signal to act on except the number itself, and the number alone does not tell you whether 680 MW is a stable estimate or a wide-error-bar median.
A practical commitment framework
A workable approach ties reserve commitment tier to the intra-day P50–P90 spread at the portfolio level:
- Low spread (<5% of nameplate): Commit against P50. Add standard spinning reserve for normal operating requirements per NERC BAL-002 minimums.
- Moderate spread (5–15% of nameplate): Commit against P65–P70. Increase fast-response reserve allocation proportionally. Flag ramp event windows for the real-time operator.
- High spread (>15% of nameplate): Commit against P85–P90. Consider holding additional non-spinning reserve available-to-dispatch. Notify interchange counterparties of elevated variability risk.
This is not a formula — it is a framework that a resource adequacy team would calibrate to their specific reserve margin, balancing area size, and thermal fleet ramp rates. The thresholds need to reflect the cost tradeoff at that balancing authority's specific unit mix.
The inertia and frequency response dimension
There is a second layer to this argument that gets insufficient attention in dispatch economics discussions: synthetic and physical inertia. As thermal generation retires and is replaced by inverter-based resources, the inertial response available to arrest a frequency deviation following a sudden generation shortfall decreases. In a high-VRE system, a 100 MW solar under-production event that would have been absorbed by spinning mass in a thermal-heavy fleet may now require a faster, larger governor response from the remaining thermal units.
This means the P90 tail matters more in a high-VRE grid than it did in a thermal-dominated one — not because the forecasts are worse, but because the consequence of a tail miss is larger when frequency recovery depends on fewer synchronous generators. Resource adequacy standards that were calibrated for a system with 80% thermal penetration may need to be re-examined when that drops to 50% or below.
P90-committed reserve in that context is not conservatism — it is a proper accounting of system vulnerability. A 10% probability tail event that causes a NERC Reliability Standard violation and emergency actions is not an acceptable operational risk just because it carries a 10% label.
What this requires from your forecast system
Implementing a P50/P90-aware commitment workflow requires that your forecast system produce calibrated probabilistic output, not just a point estimate with nominal error bars. Calibration here has a specific technical meaning: if the system claims 90% of outcomes will exceed the P90 value, that should hold empirically when you evaluate it against 12 months of out-of-sample actuals. A model that consistently over-states its own certainty — narrow confidence bands that don't reflect actual forecast error distribution — will give you P90 numbers that behave like P70 in practice.
Post-hoc calibration checks are standard practice in NWP verification. Running them on your solar forecast system before building dispatch workflows around the probabilistic outputs is not optional — it is the engineering due diligence that makes the whole commitment framework valid.
The choice between P50 and P90 is not a forecasting question. It is a reserve margin management question that happens to require accurate probabilistic forecasting as an input. Getting that input right is where the work is.