| Vose Software

Industry: Utilities
Product: ModelRisk
Application: Energy efficiency optimization under uncertainty


The Engineer Promised 3,200 MWh. The Meters Verified 2,807: Quantifying Realisation-Ratio Risk on a Retrofit Guarantee

An engineering study for a lighting, HVAC, and controls retrofit at a municipal water-treatment plant produced a confident number: 3,200 MWh/yr of electricity savings. On the strength of it, the energy-service company (ESCO) signed a performance guarantee promising 2,720 MWh/yr — 85% of the engineered figure — and agreed to write a shortfall cheque if the meters fell short. The trouble is that engineering models, measured against verified savings across thousands of retrofits, systematically over-predict. The number that actually decides whether the guarantee is met is not the engineered saving; it is the realisation ratio — the fraction of the engineered saving the meters confirm.

Rebuilt in ModelRisk, the realisation ratio comes out right-skewed with a mean of 0.88, a P10 of 0.75, and a P90 of 1.01 — centred well below the 1.00 the deterministic model silently assumes. The verified saving therefore has a mean of 2,807 MWh but only a 59% chance of clearing the 2,720 MWh guarantee. The headline distribution shows exactly where the guarantee lives.

Realisation ratio distribution

Why a point estimate fails here

The deterministic study reports 3,200 MWh and stops. Implicitly, it assumes the realisation ratio is exactly 1.00 — that every assumption about run hours, occupant behaviour, model fidelity, and post-install commissioning holds perfectly. The simulation shows that 1.00 sits near the 88th percentile of the realisation-ratio distribution: it is an optimistic outcome dressed up as the expected one. A guarantee priced off it looks free; priced off the real distribution, it carries a 41% chance of a shortfall payment. The point estimate does not just lose the spread — it anchors the contract to the wrong number.

Building the realisation ratio from the bottom up

The realisation ratio is constructed as a product of multiplicative effects, which keeps it strictly positive and gives it the right-skew the M&V literature reports (skewness ≈ +0.28 in the simulation):

  • Engineering-model bias — LogNormal centred at 0.95 with a 7% spread: models over-predict on average but occasionally the kit beats spec.
  • Operating hours / duty cycle — LogNormal around 1.00 (±8%): the equipment runs more or fewer hours than the design assumed.
  • Take-back (rebound) — a small Beta-distributed haircut: cheaper-to-run space gets used more, clawing back part of the saving.
  • Persistence — early-life commissioning faults that drag year-one savings below design.

A shared commissioning-quality factor, drawn Beta(6,2), multiplies several of these at once. This is the modelling decision that matters: a badly commissioned job is bad on rebound and persistence and run-hours together, so the downside scenarios are coherent rather than implausibly diversified away. Without that common factor, the product of near-1 terms would collapse toward its mean and understate exactly the left tail the guarantee is exposed to.

Setting the guarantee: how aggressive is too aggressive?

The guarantee level is a negotiation, and the simulation turns it into a probability curve instead of a hunch.

Probability of meeting the guarantee versus guarantee level

Sweeping the guarantee from 70% to 100% of the engineered saving shows the trade-off precisely: a guarantee at 80% of engineered (2,560 MWh) is met 76% of the time, at 85% it is 59%, at 90% it drops to 40%, and at 95% only 24%. The ESCO had signed at 85% — a level the meters confirm barely better than a coin flip. Pulling the guarantee back to the 80% point lifts the meet probability to 76% and moves the contract onto defensible ground, with the 90% planning threshold reached only below the engineered figure.

What drives the verified-saving shortfall

The guarantee is exposed to the lower tail of verified savings, so the drivers are ranked by their effect on the P10 verified saving (2,387 MWh).

Tornado of drivers of P10 verified saving

Commissioning quality is the dominant driver at ±560 MWh of P10 verified saving — more than the engineering-model bias (±300 MWh) and operating-hours uncertainty (±250 MWh) combined. This reframes the whole risk: the largest lever on whether the guarantee is met is not refining the engineering calculation but funding a proper commissioning and fault-detection regime, which compresses the left tail the contract pays out on.

The guarantee's value lives in the left tail

For the ESCO, the guarantee is an option written on the realisation ratio: pay a shortfall when verified savings fall short, earn a gain-share when they over-deliver.

ESCO net position on the savings guarantee

Across simulated M&V outcomes the ESCO's net position averages roughly break-even, with a P10 of about −$37k and a P90 near +$30k, and a 41% probability of paying a shortfall. The mean hides the asymmetry the contract creates: the downside is a real cheque in two years out of five, driven almost entirely by the realisation-ratio left tail that the deterministic 3,200 MWh never acknowledged.

What the model changed

  • The guarantee was renegotiated from 85% to 80% of engineered savings (2,560 MWh), lifting the meet probability from 59% to 76% — the simulation turned a coin-flip commitment into a defensible one.
  • A funded commissioning-and-fault-detection scope was added, justified by its ±560 MWh effect on the P10 verified saving — the single biggest lever on the left tail, ahead of any engineering refinement.
  • The ESCO priced the shortfall risk explicitly, reserving against a 41% chance of a payout rather than booking the deterministic "guaranteed" outcome as certain.
  • EM&V budget was reallocated away from re-checking the well-characterised engineering model toward measuring run-hours and persistence, the inputs that actually move the verified number.

ModelRisk functionality used

  • Realisation ratio built as a product of LogNormal and Beta factors, producing the right-skewed, sub-1.0 distribution (mean 0.88, skew +0.28) that a Normal assumption around 1.00 would have missed entirely.
  • A shared Beta(6,2) commissioning-quality factor linking rebound, persistence, and run-hours, so poorly-executed jobs are coherently bad and the left tail is not diversified away.
  • Threshold sweep of P(meet guarantee) across guarantee levels (76% / 59% / 40% / 24% at 80 / 85 / 90 / 95%), turning the guarantee negotiation into a probability curve.
  • Tornado on P10 verified saving identifying commissioning quality (±560 MWh) as the dominant lower-tail driver, ahead of engineering-model bias.
  • Custom guarantee-and-gain-share logic evaluated per simulated outcome to produce the ESCO's net-position distribution and the 41% shortfall-payment probability.

In efficiency retrofits, the engineered saving is the question, not the answer — the realisation ratio is the answer, and Monte Carlo is what puts a number on it before the guarantee is signed rather than after the first reconciliation.