Windgrams

← Research

Entry 07Ensemble interpretation3 min read

What ensemble spread can—and cannot—tell you

A deterministic windgram draws the same crisp line on an easy forecast and a sensitive one. REPS runs 21 related atmospheres from perturbed initial conditions. Agreement shows that those model realizations respond similarly; separation shows that one deterministic line hides sensitivity.

The ensemble publishes four scalar distributions

REPS exposes only three pressure levels through the displayed windgram band. At the catalogued mountain launches, the lowest is often below model terrain elevation. Twenty-one full windgrams drawn from those sparse columns would add authority without adding vertical evidence.

The ensemble product therefore publishes member spread for cloud base, boundary-layer top, thermal velocity, and usable-lift top. It does not publish 21 atmospheric profiles. The forecast model feed reference holds the current level inventory and provider paths.

Calculate percentiles from derived member values

Each member passes through the complete nonlinear derivation as its own atmosphere. Percentiles are calculated across the resulting scalar values. Averaging temperature, moisture, and heat flux first would manufacture an atmosphere no member predicted, and its boundary-layer top would not equal the average of the members’ tops.

Per forecast hour and scalar, the schema is:

"usableLiftTopM": {
  "ceiledMembers": 0,
  "members": 21,
  "p10": 3222.3,
  "p25": 3355.1,
  "p50": 3503.9,
  "p75": 3530.2,
  "p90": 3647.3
}

members separates occurrence from magnitude: a member with no usable lift does not enter the height ranking. ceiledMembers counts values clamped at the top of the available column, where the reported height is a lower bound. Wind direction has no percentiles because circular values cannot be ranked linearly: the midpoint of 350° and 10° is not 180°.

Count occurrence across 21 members before ranking height

Choose the accounting step
  1. M01
  2. M02
  3. M03
  4. M04
  5. M05
  6. M06
  7. M07
  8. M08
  9. M09
  10. M10
  11. M11
  12. M12
  13. M13
  14. M14
  15. M15
  16. M16
  17. M17
  18. M18
  19. M19
  20. M20
  21. M21
18 / 21members return a usable-lift top

The other 3 return null because their derived core never clears the sink threshold. Null is absence of the quantity, not a zero-metre height, so those members do not enter the height ranking.

Illustrative 21-member values constructed to show schema semantics. The member counts and percentile arithmetic are literal; the elevations are not an archived forecast case.

The column ceiling and 500 hPa

One strong forecast hour placed all 21 members at the 700 hPa column ceiling, about 3,150 m. Their boundary-layer tops differed by only 16 m. The narrow spread described the shared input ceiling, not the weather.

Adding 500 hPa as computational headroom raised the ceiling to about 5,700 m. For the same hour, the boundary-layer spread widened from 16 m to 205 m and usable-lift spread widened to 425 m. Hours that had already terminated below 700 hPa remained bit-for-bit unchanged. ceiledMembers preserves the warning if a later atmosphere reaches the higher ceiling.

A short pressure column censors ensemble spread

Three pressure levels

column top ≈ 3,150 m
SOURCE DATA END HEREapparent spread ≈ 16 mMORNINGAFTERNOON

Censored. Every trace stops at the highest supplied level. The tight result measures the input ceiling.

Column extended to 500 hPa

headroom above 3,150 m
FORMER CEILINGresolved member rangeMORNINGAFTERNOON

Resolved. With vertical headroom, the same members separate. The range can now describe model disagreement.

Nine representative traces from the 21-member ensemble. Shapes are illustrative; the short-column ceiling and the added 500 hPa level reproduce the censoring test.

Read spread in this order

  1. Member fraction: “18 of 21 members produce usable lift” comes before any height band.
  2. Median: p50 is the ensemble centre, not the deterministic HRDPS line.
  3. Conditional range: p10–p90 describes members that produced the quantity.
  4. Ceiling count: a nonzero count makes the upper statistics lower-bound evidence.

Do not draw the REPS range as an error bar around HRDPS. Different model terrain elevations and physics can put a deterministic line outside the ensemble range on an ordinary day.

Spread is not probability

Member fraction measures agreement among model realizations. “18 of 21” does not establish an 86% chance of usable lift. Probability requires calibration against archived forecasts and observations; this project does not yet have that evidence.

Provider and transport details live in the forecast model feed reference. ECCC documents the underlying system on its REPS open-data page.