Reference document — available throughout all three days. Print this page or keep it open.
Building Profile
12,000 sf single-story municipal building
ASHRAE Climate Zone 4A (Mid-Atlantic — Washington Dulles)
Gas heat + DX cooling · Two packaged rooftop units
One electric meter · One gas meter
24-month baseline · 12-month reporting period
The CMVP serves the agreement. The ESCO wants conservative baselines. The municipality wants aggressive savings claims. The CMVP facilitates agreement before the baseline is built — then executes the plan both parties signed.
Project Timeline
Baseline Period
24 months of pre-retrofit data
Install
ECM-1 & 2
Reporting Period
12 months post-retrofit
Month 9: NRA
NRA Event — Month 9: Community room addition (1,200 sf) with new HVAC zone. Known in advance. Pre-documented in M&V plan with agreed adjustment methodology. Adds ~420 kWh/month to electric consumption.
Utility Rate Structure
Electric (blended)
$0.105 / kWh
Natural Gas
$1.20 / therm
Est. Annual Cost
$11,701 elec + $9,109 gas = $20,810
Demand Charge
Not modeled in this capstone (simplification)
Teaching point (Day 3): The blended rate ($0.105/kWh) is a simplification. Marginal rates vary by tier and season. The difference between blended and marginal valuation is quantified in the cost-avoidance calculator.
The IDF is the energy audit in computable form.
The EnergyPlus Input Description File encodes every system, schedule, load, and assumption about this building as-found. It is not just a simulation input — it is the documented epistemological basis for every boundary decision, every model assumption, and every Non-Routine Adjustment trigger that follows.
A rigorous energy audit should produce an IDF. Everything else — the regression model, the savings estimate, the uncertainty budget — is downstream of that record.
Sections 1 · 2 · Cuts are embedded where the section naturally pauses
§ 1
Fundamental Concepts
The counterfactual method · Baseline vs. counterfactual · Types of models (statistical, physical, hybrid) · Professional judgment · Uncertainty · The role of the CMVP
75 slides
▶ CUT 1
Boundary Setting
↳ After "Three Types of Models" — "Where does the measurement stop?"
Boundary Decisions — ECM-1 LED Retrofit
1. ECM-1 lighting circuit (office wing, Panel A)
2. HVAC response to lighting heat gain change
3. Plug loads and other tenant equipment
Who turns off the lights? If building staff control operating hours manually, hours are an operational factor — the owner bears the risk of variation. If automated controls drive hours, it's a performance factor — the contractor is responsible. Reference: ASHRAE Guideline 14 §5 | FEMP M&V §3.2
📖
Kromer (2024) — Chapters 1 & 2
The counterfactual method · Energy management · The denominator
"Energy management involves applying the fundamental laws of physics to energy systems with a goal of optimizing the value produced per unit of energy consumed."
The boundary decision is made once and governs everything downstream. ECM-1 (lighting) can be isolated at the office wing subpanel — the physics allows it. ECM-2 (controls) cannot be isolated from the whole-facility meter — the savings are invisible at the equipment level. This distinction forces two different approaches for one building.
FEMP M&V Guidelines 4.0 — single-line diagrams and meter placement
Judea Pearl — The Book of Why(counterfactual reasoning foundations)
Stewart Brand — How Buildings Learn(buildings evolve; baselines must account for this)
§ 2
Contextual Considerations
Four risk domains (technical, regulatory, commercial, legal) · ESPC and utility programs · Stakeholders · Guiding principles · Cost and risk · Guidelines and protocols
44 slides
▶ CUT 2
Model Form Selection
↳ After "Guiding Principles" — "Which model for which ECM?"
Decision Tree — Select model form for each fuel
Electric Model
Q1: Does the building have significant cooling load driven by outdoor air temperature?
Gas Model
Q2: Is heating load the primary driver of gas consumption?
Interactive Effects
Q3: Do the ECMs affect electric and gas independently?
Review your answers above — one or more selections need to be revisited before proceeding.
CFHQ Model Selection — Confirmed
Electric: 3PC
Cooling-only change-point. Baseload + cooling ramp above OAT threshold. No heating ramp in electric data (gas heat).
Gas: 3PH
Heating-only change-point. Gas consumption rises below OAT threshold. Summer baseload near zero (DHW only).
Interactive effect: LED retrofit removes heat gain → gas heating load increases in winter. The models are separate but the savings are coupled. This must be documented in the M&V plan.
📖
Kromer (2024) — Chapters 3 & 4
Approach selection · Retrofit isolation vs. whole facility
"Option letters (A/B/C/D) are not the answer. The physics of the ECM is the answer."
Model form selection is a physics decision, not a statistics decision. LED wattage doesn't drift — stipulation is defensible. RTU controls savings are distributed across the whole building — only a whole-facility inverse model can capture them. The student who understands why each ECM demands a different approach has learned more than the one who memorizes the Options table.
Key References
CfD Framework: Retrofit isolation (ECM boundary) vs. whole-facility (regression boundary) — a Boundary decision, not an Option letter
FEMP 4.0 — recommended M&V approaches by ECM type (Table 3-1)
ASHRAE Guideline 14 — performance verification vs. counterfactual method
ISO 50015:2014 — M&V of energy performance of organizations
↳ After sampling slides — "What do you see in this data?"
Electric vs. OAT · Gas vs. OAT · Raw points only, no model · Students observe and write before any fitting
📖
Kromer (2024) — Practical Considerations
Descriptive statistics · Sampling uncertainty
"Ten motors sampled from a production floor returned a CV of 8.38% — a real result from real variation, not model error."
Scatter plots before modeling is a pedagogical choice. Students who see the shape first develop physical intuition — they understand why the 5P model has two change-points before they ever move a slider.
BPA Uncertainty Guide — confidence levels and precision for M&V
§ 4
Baseline & Adjustments
Three baseline concepts · Baseline period · Static factors · Routine adjustments · Non-routine adjustments · Avoided vs. normalized vs. backcasting
23 slides
▶ CUT 4
Baseline Period
↳ After "Three Baseline Concepts" slide — before static factors discussion
24-month time series · Students select 12-month window · Stage 0B service-level check · Static factor inventory for CFHQ · Pre-document NRA trigger
📖
Kromer (2024) — Chapters 1 & 2
Energy management · Lifecycle · The denominator · Service levels
"The only fair and relevant baseline is one where the facility meets all codes, regulations, and service requirements."
The three baseline concepts — period, data, and model — must all be specified in the M&V plan. Most disputes trace back to ambiguity in one of the three. The denominator question (energy per what?) comes from Chapters 1–2 and underlies every baseline period judgment.
Key References
FEMP M&V Guidelines 4.0 — baseline period selection criteria
ASHRAE Guideline 14 — 90% coverage factor requirement for baseline period
IPMVP Core Concepts — static factor documentation and NRA triggers
NIST Life Cycle Costing — RUL, EUL, dual baselines (referenced in § 1 slides)
⚡
MCP Demo — Forward Model
At the § 4 → § 5 transition · Claude Desktop + EnergyPlus-MCP · ~25 minutes
"This is the answer key. Now forget it. The rest of this session, you work from the meter data backward."
↳ After Fractional Savings Uncertainty slides — "Now fit the inverse model"
Scatter plot · Change-point sliders · Live R², CV(RMSE), NMBE · Class spread shows inter-analyst uncertainty · Compare to EnergyPlus savings estimate
📖
Kromer (2024) — Practical Considerations
M, M & V — measuring, modeling, and verification
"The model is the most important element of M&V. Without a defensible model, savings are not savings — they are assertions."
The inverse model doesn't measure savings — it infers them. The baseline relationship learned from historical data is projected forward into the reporting period. That projection is the counterfactual: what would have happened. It is a designed artifact, not a measurement. The spread of change-point choices across the class is not error — it is the honest range of defensible professional judgment.
Key References
Reddy & Claridge — "Uncertainty of 'Measured' Energy Savings from Statistical Baseline Models" (FSU methodology)
CalTrack / NMEC — whole-facility statistical approach, utility program context
Touzani et al. — "Evaluation of Methods to Assess Uncertainty in Estimated Energy Savings"
§ 6
Retrofit Isolation
Single-line diagrams · Measurement boundaries · Performance vs. performance/operation · Interactive effects · Significant vs. estimable parameters
22 slides
▶ CUT 6
Boundary + Stipulation
↳ After terminology slide — "Draw the measurement boundary"
Annotate single-line diagram · ECM-1 office panel vs. ECM-2 main meter · Lighting stipulation: CV from sample → required N · Boundary as risk allocation decision
Dimension 1 — Cost
Retrofit isolation costs more upfront. It requires sub-metering or dedicated instrumentation, more M&V plan design time, and more analyst hours per reporting period. But it transfers risk — the party paying for M&V is buying certainty about one thing: did the ECM perform as specified? Everything else (occupancy, plug loads, process changes) is outside the boundary. The verified claim is narrow but defensible.
Whole-facility is cheaper — but the savings claim absorbs everything that happened in the building, good and bad. Lower cost, higher exposure.
Dimension 2 — Control & Responsibility
Who controls what? This is the harder question — and the one that determines who bears the risk when numbers don't match expectations.
Performance factors (within ECM boundary)
Equipment efficiency (watts/sf, COP, EER)
Controls logic (setpoints, schedules)
Installation quality
→ The contractor/vendor controls these
→ Retrofit isolation verifies these
Operational factors (outside ECM boundary)
Occupancy hours and density
Plug loads and process equipment
Thermostat overrides by building staff
Building additions or renovations
→ The building owner/operator controls these
→ Whole-facility absorbs these
The boundary decision is really a question of who bears the risk of operational variation. If the owner changes occupancy hours after installation and savings drop — who is responsible? Under retrofit isolation, the owner bears that risk (it's outside the boundary). Under whole-facility, it shows up as reduced savings, and the contractor may be blamed unfairly.
Retrofit Isolation
Whole Facility
Boundary
ECM only
Whole meter
Cost
Higher
Lower
Claim
Narrow, defensible
Broad, exposed
Performance risk
Contractor
Contractor
Operations risk
Owner
Shared / disputed
Best for
ESPCs, performance contracts
Portfolio screening
Who turns off the lights?
That answer determines whether operating hours are a performance factor or an operational factor — and who bears the risk if they change. Is it an automated schedule (controls-driven = performance factor, contractor's responsibility)? Or building staff flipping switches (behavior-driven = operational factor, owner's risk)? Or both — automated schedule with manual override? If it's manual or mixed, operating hours are an operational factor. This is why the stakeholder and risk matrix (Step 1 in the capstone flow, pre-Cut 1) is not administrative overhead — it's the document that determines who is liable for what when the numbers don't match expectations.
"Stipulation is not weakness — it is an explicit, auditable assumption. The question is not whether to stipulate, but whether the stipulation is defensible."
Performance vs. performance/operation is a cost-precision tradeoff, not a quality ranking. Stipulated-parameter measurement can be more defensible than continuous metering if the stipulated parameter is genuinely stable. The CV of the sample drives the required N — students who calculate this by hand understand why sampling plans aren't arbitrary.
Stipulation is a contractual decision, not a shortcut. When you stipulate an operational factor — say, hours of use — you are explicitly assigning the risk of variation in that factor to the owner. If actual operating hours drop 20% after installation because the tenant changed their schedule, the contractor's lighting ECM performed exactly as specified. The savings shortfall is an operational variance, not a performance failure. The M&V plan records that assignment of risk. The stipulation is the mechanism.
The stakeholder map must answer "who controls operating hours?" before the M&V plan is written — not after a dispute arises. For the CFHQ lighting ECM: if the office panel schedule is automated by the BAS, operating hours are a controls-driven performance factor under the contractor's scope. If building staff routinely override the schedule, those hours become behavior-driven and belong in the operational column. Mixed control (automated with manual override) means the stipulated value carries real risk — and the owner has formally accepted it.
Sections 7 · 8 · 9 · Cuts are embedded where the section naturally pauses
§ 7
Reporting Period Analysis
Applying the counterfactual · Monthly savings · Interactive effects · Gas increase explanation · Savings bar chart
18 slides
▶ CUT 7
ECM-1 Savings Calculator
↳ After "Adjusted Baseline Model" review — "Now apply your model to the reporting period"
Panel A — Apply the Baseline Model
The baseline regression model (from Cut 5) predicts what energy consumption would have been at each month's actual outdoor temperature. The difference between predicted and actual is the inferred savings — not measured, inferred.
Month
OAT (°F)
Model-Predicted Baseline
Actual Post-Retrofit
Inferred Savings
The baseline model is applied to reporting-period conditions — not copied from the baseline period. This is the counterfactual: what would consumption have been, given this month's weather, had the ECM not been implemented?
Interactive effect: ECM-1 lighting retrofit REDUCES electric consumption but INCREASES gas consumption in winter — lighting heat gain was part of the heating load. The old T8 fixtures warmed the office wing; the new LEDs produce less heat. The HVAC must compensate. This is expected, quantifiable, and must be reported.
Panel C — Explain Your Answer
In your own words: why does ECM-1 increase gas consumption in winter? What does this tell you about boundary setting?
"The counterfactual makes the gas increase legible. Without the model, the building manager sees a higher gas bill and concludes the project failed."
Interactive effects are not errors — they are physics. The LED retrofit removed internal heat gain from the lighting fixtures. In winter, the HVAC must now compensate for the lost heat. Gas consumption rises. The model predicted this. The M&V plan documented it. The student who can explain this to a non-technical stakeholder has mastered the reporting period.
↳ After "Static Factors" review — "Document what you expect BEFORE you look at the data"
Non-Routine Adjustments must be pre-documented BEFORE the event occurs. Reverse-engineering an NRA from a residual spike is not M&V — it is storytelling. The pre-documentation below anchors your adjustment to a prior, falsifiable prediction.
Known Non-Routine Factors (from CFHQ IDF)
Zone / System
Variable
In Baseline Model?
NRA Risk
Building Envelope
Gross floor area (12,000 sf)
Yes
—
Building Envelope
Window area / SHGC
Yes
—
OFFICE
Occupancy density (5/1000 sf)
No
Medium
LOBBY
Occupancy density (10/1000 sf)
No
Medium
OFFICE
Plug load density (1.0 W/sf)
No
High
LOBBY
Plug load density (0.5 W/sf)
No
Medium
OFFICE
Operating schedule (weekday 7a-6p)
No
Medium
LOBBY
Operating schedule (7a-9p extended)
No
Medium
HVAC — Both RTUs
Cooling COP (3.50)
Yes
—
HVAC — Both RTUs
Gas furnace efficiency (0.80)
Yes
—
Building Envelope
Infiltration rate (0.06 cfm/sf EWA)
No
Low
Building Geometry
Floor area addition / renovation
No
High
HVAC — Both RTUs
Equipment replacement or capacity change
No
High
Factors outside the baseline model boundary are NRA candidates. If they change during the reporting period, a pre-documented adjustment is required before the event occurs.
NRA Pre-Documentation Form
✓ NRA pre-documented. Your estimate will appear as a reference line in Cut 9.
📖
Kromer (2024) — Chapters 5 & 6
Non-routine adjustments · Static factors · The M&V plan as contract
"Non-routine adjustments are not surprises — they are anticipated in a well-written M&V plan."
Pre-documentation is the discipline that separates planning from post-hoc rationalization. The community room expansion at CFHQ was known before the performance period began. A well-written M&V plan includes it as a planned NRA with a pre-agreed adjustment methodology. Failing to document it in advance is a planning failure, not a reporting surprise.
Capstone flow — Stage 7: Static Factors and NRA Protocol
§ 9
NRA Discovery
Residual analysis · Step change detection · Month 9 community room addition · Pre-doc vs. discovery comparison · The payoff moment
16 slides
▶ CUT 9
Residual Plot Tool
↳ After "Detecting Non-Routine Events" — "Now look at the residuals"
Panel A — Reporting-Period Residuals · Electric only — gas residuals not shown in this view
Each bar shows the difference between the baseline model prediction and actual post-retrofit consumption. Months near zero = the model explains the data. A persistent step change = something outside the model changed.
Complete Cut 8 to overlay your pre-documented NRA estimate.
Panel B — Apply Non-Routine Adjustment
NRA adjustment: OFF
✓ Adjustment accepted. Inferred savings are valid.
Months 9–12 show a persistent positive residual. The model is under-predicting consumption. Something changed.
The NRA does not change what happened — it changes what the baseline is expected to have been. The counterfactual shifts. The inference holds.
Panel C — Goodness-of-Fit: Reporting Period Diagnostics
Metric
Months 1–8
Months 9–12 (unadjusted)
CV(RMSE)
3.2%
18.4% (fails)
Mean Residual
-28 kWh
+387 kWh
ASHRAE 14 / IPMVP require CV(RMSE) ≤ 15% monthly for whole-facility models. A persistent residual pattern is a diagnostic signal — not a model failure. The NRA restores model validity.
Panel D — Explain Your Answer
In your own words: what would happen to the inferred ECM-1 savings if this NRA were NOT applied? Who bears the risk?
✓ Reflection submitted.
📖
Kromer (2024) — Practical Considerations
NRA discovery · Residual analysis · The counterfactual payoff
"Without the NRA, the controls upgrade appears to have failed. With it, savings are real and positive. The model didn't fail — the building changed."
This is the payoff moment. The student who pre-documented the community room expansion in Cut 8 now sees it appear in the residuals — exactly where they predicted it. The student who didn't pre-document it sees a mysterious step change and must explain it after the fact. The difference is the difference between M&V planning and M&V forensics. The former is professional; the latter is expensive.
You have 12 months of baseline data for Counterfactual Headquarters. You just watched EnergyPlus run the forward model.
Now work from the meter data backward. Fit a statistical baseline model — the inverse direction.
The Decision
Where are the change-points?
Statistics tell you fit. R² and CV(RMSE) tell you how well the model describes the data.
Physics tells you sense. Does 38°F as a heating threshold make sense for a Zone 4A building with gas heat?
The class will produce a spread of change-point choices — all defensible. That spread is inter-analyst uncertainty. It's not error. It's honest.
Explainer
This is the inverse model — the statistical counterfactual machine. It learns the baseline relationship between weather and energy use, then projects that relationship into the reporting period.
Savings = Counterfactual − Actual.
The model is the most important element.
Navigation
Baseline Model — Counterfactual Headquarters
Model form:← this data has no clear heating ramp; 3PC fits better
R²
—
CV(RMSE)
—
NMBE
—
vs. EnergyPlus Forward Model — ECM-1 · LED Lighting
EnergyPlus predicted annual savings— kWh
Your inverse model estimate— kWh
Difference—
The gap between forward and inverse model estimates is model uncertainty — not error. Both are defensible representations of a building we cannot observe directly.