Dr. Eshan SinghGrid Reliability Lab Student guide Open the lab
Resource adequacy · from first principles

How planners decide whether a grid has enough resources

Reliability planning is not just checking whether installed megawatts exceed peak demand. It asks whether a changing portfolio can serve every hour across uncertain weather, equipment failures, fuel conditions, imports, storage state of charge, and demand.

By the end, you should be able to
  • Separate adequacy from operating and distribution reliability
  • Interpret LOLE, LOLH, LOLP, LOLEV, EUE, NEUE, and annual risk
  • Explain weather years, outage draws, and Monte Carlo uncertainty
  • Describe firm need, EFC, and ELCC without confusing them
  • Challenge a reliability result with industry-style questions
01 · The planning question

Reliability has several meanings

This lab focuses on bulk-system resource adequacy: whether the system has enough deliverable capacity and energy to serve firm demand over time, allowing for uncertainty.

Resource adequacy

Planning for enough generation, storage, demand response, and imports to meet demand across plausible conditions.

Modeled here

Operating reliability

Keeping frequency, voltage, reserves, and transmission secure in real time after credible disturbances.

Not solved here

Distribution reliability

Customer interruption frequency and duration on local wires, often summarized with SAIDI and SAIFI.

Not solved here
A useful mental model

Resource adequacy is an hourly balance repeated across many plausible years

A planning reserve margin compares a single peak-demand number with an adjusted capacity total. It is useful for screening, but it cannot see whether solar output fades before the peak, a battery empties during a long event, wind is correlated across regions, or several generators fail together. Probabilistic adequacy replaces that single comparison with thousands of chronological hourly balances.

Available supplyafter outages and derates+Discharge + demand responsewithin power and energy limitsFirm demandweather and forecast dependent=Hourly marginnegative means shortfall

The model records every negative margin, then aggregates the same events in several ways. That is why LOLE, LOLH, LOLEV, and EUE are complementary rather than competing answers.

02 · The physical event

A shortfall occurs when available supply cannot cover firm demand

For each hour, the model compares firm demand with the supply that is actually available after renewable profiles, outages, derates, import limits, demand response, and storage dispatch. Any remaining gap is unserved energy.

Illustrative stressed day

One event can create several different metrics

Firm demandAvailable supplyShortfall
Shortfall hourshourly demand points above supply
Peak deficitlargest hourly demand–supply gap
Unserved energysum of hourly deficits × 1 hour
Every red boundary is calculated from the same hourly demand and supply arrays as the two lines. At a crossing, the chart interpolates the exact intersection; the fill cannot drift away from either curve.
How to read it

Width, height, and area say different things

The horizontal span tells you when and how long the deficit lasts. The vertical separation tells you how deep the deficit is. Their hourly sum is EUE. A thin, deep event and a broad, shallow event can have the same EUE but different operating consequences.

Worked arithmetic

Five hourly deficits become 39 GWh

The plotted deficits are 6, 10, 12, 9, and 2 GW. With one-hour intervals, EUE is (6 + 10 + 12 + 9 + 2) GW × 1 h = 39 GWh. Because all five hours occur on one calendar day and form one uninterrupted episode, that case-year contributes 1 LOLE day, 5 LOLH hours, and 1 LOLEV event.

Do not infer

Adequacy shortfall is not a customer-outage forecast

The red region is system-level energy not served before emergency actions. It does not identify which customers would be interrupted, for how long, or whether transmission and distribution equipment can deliver available generation.

1Demand risesHeat, electrification, or forecast error can increase load.
2Supply changesOutages, weather, imports, hydro, and maintenance affect availability.
3Flexibility respondsDemand response and storage reduce the deficit within their constraints.
4The gap is countedOccurrence, duration, depth, and event count become different metrics.
03 · Building evidence

One historical year is not enough

Rare-event planning needs many chronological case-years. This lab pairs each weather chronology with independent year-long equipment failure and repair sequences. Every case preserves all 8,760 hours so storage, event duration, and seasonal patterns remain meaningful.

25weather years
×
50outage draws / year
=
1,250simulated case-years

Weather years

Different synchronized hourly patterns of demand, solar, and wind. They represent the finite range of weather chronologies in the input library.

Outage draws

Different random sequences of equipment failures and repairs applied to each weather year. More draws reduce Monte Carlo sampling noise; they do not create new weather.

Chronology

Hour 4 depends on what happened in hours 1–3

A chronological model carries state forward. A failed unit remains unavailable until it repairs. Storage state of charge rises when it charges and falls when it discharges. Demand-response energy budgets reset on their stated schedule. Shuffling the same 8,760 hours could therefore change the answer.

Correlation

Weather variables must stay synchronized

A hot evening may combine high load, weak wind, reduced thermal output, and constrained imports. Sampling each input independently can create impossible combinations or erase genuine compound stress. The California ensemble keeps load, solar, and wind aligned within each weather chronology.

Weighting

Case-years do not automatically have equal meaning

Within this lab, each weather year receives the same number of outage draws and weather years are equally weighted. A production study may use explicit weather weights, climate-adjusted futures, multiple load forecasts, or scenario probabilities. The reported mean must reflect those weights.

Conditioning

The answer is conditional on the modeled world

“Expected” means expected across the study’s weather records, outage processes, dispatch rules, and input assumptions. It does not mean every uncertainty in the real grid has been sampled. Fuel constraints, transmission topology, hydro chronology, and emergency procedures may require separate modeling.

Facility and data-center studies

Demand shape and supply weather are separate choices

A data center may draw nearly constant power even though the surrounding grid has a strongly seasonal and daily load shape. Applying California system demand to that facility would answer the wrong question. In the lab, choose Constant 24/7 or upload the facility’s own 8,760-hour load profile, then independently choose the 25-year supply-weather set.

What stays fixed

The facility load shape is repeated across every selected supply-weather year. A flat 100 MW case remains 100 MW in every hour before the optional annual forecast multiplier and cooling uplift. The model does not reshape it into California system demand.

What still changes

Solar and wind output, weather-linked thermal and import stress, hydro uncertainty, common-mode events, and equipment failures still vary. This tests the same customer demand against many plausible bulk-supply conditions.

04 · The industry vocabulary

Each metric answers a different question

Let Ui,h be unserved demand in hour h of simulated case-year i. An indicator 1(·) equals one when its condition is true and zero otherwise.

LOLEdays/year

On how many calendar days does at least one shortfall hour occur, on average?

meani Σdays d 1(any Ui,h in d > 0)
LOLHhours/year

For how many hourly intervals is firm demand not fully served, on average?

meani Σh 1(Ui,h > 0)
LOLEVevents/year

How many separate contiguous shortfall episodes occur, on average?

meani count(event starts)
EUEMWh/year

How much firm energy is not served, capturing both depth and duration?

meani Σh Ui,h × 1 hour
NEUEppm

What share of annual demand energy is expected to be unserved?

EUE / annual demand energy × 106
Hourly LOLPprobability

What fraction of simulated cases has a shortfall in a particular hour?

meani 1(Ui,h > 0)
Annual risk% of years

What fraction of simulated years contains at least one shortfall hour?

meani 1(any Ui,h > 0)
MetricOccurrence?Duration?Depth?Common mistake
LOLEYes, by dayNot within a dayNoTreating 0.1 day/year as 2.4 hours/year
LOLHYes, by hourYesNoCalling it LOLE without stating hours
LOLEVYes, by eventNoNoAssuming every event lasts one hour or one day
EUEIndirectlyYesYesIgnoring a few very severe tail cases
Annual riskYes, by yearNoNoEquating it with LOLE days/year
One sample, seven different summaries
Suppose 1,000 simulated case-years contain 100 shortfall-days, 180 shortfall-hours, 72 contiguous events, and 12,000 MWh of unserved energy. Sixty of those case-years contain at least one event.
LOLE = 0.100100 days ÷ 1,000 case-years
LOLH = 0.180180 hours ÷ 1,000 case-years
LOLEV = 0.07272 events ÷ 1,000 case-years
EUE = 12 MWh/yr12,000 MWh ÷ 1,000 case-years
Annual risk = 6%60 affected years ÷ 1,000 case-years

None of these values can be converted into another without the underlying event chronology and depths.

05 · Statistical discipline

A reliability estimate needs an uncertainty statement

Shortfalls are deliberately rare in a well-planned grid, so estimates can move when more outage draws are added. A point estimate without its sample count, weather basis, and confidence interval can look more precise than it is.

Illustrative 95% intervals

The interval—not the dot—determines the evidence category

0.1 d/yr criterion
These are teaching examples, not results from the current model. A confidence interval describes Monte Carlo sampling uncertainty conditional on the modeled weather set and assumptions; it is not a complete range of real-world uncertainty.
Decision logic

Compare the whole interval with the criterion

If the upper bound is below 0.1, the sampled evidence supports “below target.” If the lower bound is above 0.1, it supports “above target.” If the interval crosses 0.1, the honest conclusion is sampling-uncertain, even when the mean is on one side.

What more draws do

They reduce one uncertainty, not every uncertainty

Additional outage draws usually narrow Monte Carlo uncertainty. They do not add new heat waves, change incorrect outage assumptions, represent transmission constraints, or eliminate structural model error. Those require more weather data, sensitivities, and better model detail.

Monte Carlo uncertainty

Comes from using a finite number of outage draws. Increase draws to reduce sampling noise, particularly near a planning criterion.

Model and weather uncertainty

Comes from assumptions, data, future climate, portfolio detail, imports, hydro, and the finite weather library. More outage draws do not remove it.

06 · Planning to a criterion

Firm need translates risk into a capacity question

The firm-need study asks how much perfectly reliable, always-available capacity would need to be added—or could be removed—before the portfolio reaches a chosen LOLE target.

01Run the portfolio

Estimate its current LOLE using fixed weather years, outage draws, and seed.

02Bracket the target

Add or remove a perfect-firm proxy until results lie on both sides of the target.

03Narrow the range

Use bisection until the capacity bracket reaches the selected MW tolerance.

04Report honestly

Show the bracket, estimate, sampling interval, and whether the conclusion is statistically clear.

Firm-need conditionLOLE(portfolio + firm adjustment) ≈ selected LOLE target
Illustrative bracket7.3 to 7.4 GW

If 7.3 GW of perfect firm capacity leaves mean LOLE above 0.1 d/yr and 7.4 GW brings it to or below 0.1 d/yr, the solution is a 100 MW bracket. Reporting 7.36 GW would claim precision the search and finite Monte Carlo sample do not support. A conservative planning value uses the safe endpoint, while the midpoint is a convenient estimate.

A positive answer means perfectly firm capacity to add. A negative answer means perfect-firm-equivalent surplus that could be removed in this simplified single-area model. It is not automatically a procurement recommendation.

07 · Comparing unlike resources

EFC and ELCC measure capacity value in different ways

Capacity value asks how much reliability contribution a resource provides—not how many nameplate megawatts it has. This matters most for weather-dependent and energy-limited resources.

EFC

Equivalent Firm Capacity

Replace the candidate resource with perfectly reliable firm capacity and find the amount that produces the same risk. This lab divides EFC by candidate nameplate for its firm-equivalent capacity-credit percentage.

Risk(B + F, L) ≈ Risk(B + C, L)
B = portfolio without candidate; C = candidate; F = perfect firm capacity; L = load.
ELCC

Effective Load-Carrying Capability

Add the candidate, then find how much extra load the full portfolio can carry while returning to the base portfolio’s risk.

Risk(B + C, L + ΔL) ≈ Risk(B, L)
In this lab, ΔL proportionally scales the complete load shape and is reported as a reference-peak load increase.
Risk-matched capacity value

ELCC is a horizontal shift at equal reliability

Base portfolio BPortfolio B + candidate C
Base load at matched riskportfolio B at 0.1 d/yr
Full-system loadB + candidate C at the same risk
Illustrative ELCChorizontal difference, not nameplate
Both points are computed from the plotted equations at exactly 0.1 shortfall-day/year. Their vertical coordinates are identical; only their system-load coordinates differ.
Step 1

Define the base and candidate systems

B is the entire portfolio with the selected candidate removed. B + C restores that candidate. This definition matters: ELCC is marginal to the portfolio and can change when the rest of the system changes.

Step 2

Hold reliability risk constant

In the teaching curve, B serves 66.0 GW at 0.1 d/yr. B + C serves 68.4 GW at that same risk. The vertical matching rule prevents us from confusing lower risk at unchanged load with load-carrying capability.

Step 3

Subtract matched loads

ELCC = 68.4 − 66.0 = 2.4 GW. This is a reference-peak load increase, not a claim about instantaneous output. The lab reports EFC divided by nameplate as its firm-equivalent capacity-credit percentage.

How this lab calculates ELCC
  1. Remove the selected resource from the portfolio to form base system B.
  2. Run B at the current peak load to obtain the reference risk.
  3. Restore the candidate C, then increase load and rerun the same weather/outage sample design.
  4. Bracket the load increase whose selected risk metric—LOLE or EUE—matches the base risk.
  5. Report the ELCC bracket midpoint as a reference-peak load increase and firm-equivalent capacity credit = EFC / candidate nameplate.

Common seeds reduce comparison noise, but finite Monte Carlo estimates are stepwise. The app therefore reports a bounded estimate at the selected MW tolerance instead of false precision.

08 · Energy-limited resources

A battery’s ELCC depends on when risk occurs

A battery contributes most when it has enough power, duration, stored energy, and availability to discharge during the system’s risk hours. Adding battery nameplate does not guarantee one-for-one ELCC.

Illustrative net-load day

Charging and discharging move energy into the risk window

Before batteryAfter batteryBattery action
Original peakhighest before-battery net load
Post-battery peakhighest after-battery net load
Energy dischargedsum of discharge MW × 1 hour
The after-battery line is calculated hour by hour as before-battery net load + signed battery power. Positive power means charging; negative power means discharging. The colored areas therefore terminate exactly on both data lines.
Worked schedule

A 10 GW / 40 GWh battery shifts energy, not time

The example charges 39 GWh during lower-load hours and returns 35 GWh in the evening—about 90% round-trip efficiency. It reduces the daily peak from 68 to 58 GW. The lower panel exposes the exact signed power schedule rather than asking an area annotation to do two jobs.

Reliability interpretation

Peak reduction is not automatically ELCC

A 10 GW peak reduction on this one deterministic day does not establish 10 GW of ELCC. ELCC must be risk-matched across all weather years and outage draws. A battery may begin an event partly charged, experience an outage, face a longer event, or encounter multiple tight days without enough recharge.

DurationA longer event may exhaust a short-duration battery before the risk window ends.
Charging opportunityEnergy must be available before the event; tight multi-day systems may not recharge fully.
Risk timingSolar shifts net-load risk toward evening, changing when storage is most valuable.
Portfolio saturationAs more batteries are added, they may all serve the same first few peak hours, reducing marginal ELCC.
Outages and efficiencyUnavailable power and charging losses reduce what reaches the system during stress.
Boundary assumptionsThis lab’s storage ELCC is conditional on entered Jan 1 SOC and reliability-directed dispatch.
09 · From number to decision

Read a study as a connected story

1

Start with the study basis

Portfolio, peak demand, weather years, outage draws per year, stress mode, seed, and storage starting SOC.

2

Read occurrence and severity together

LOLE and LOLH show frequency/duration; EUE and event tables show depth and tail severity.

3

Find when risk happens

Use the month-by-hour heatmap, worst-day trace, and per-weather-year results.

4

Check sampling confidence

Inspect the LOLE interval and convergence. Increase draws when conclusions are close to the target.

5

Separate diagnostics from causes

Stress signatures show conditions present during shortfalls; they do not prove causal attribution.

6

State the model boundary

Single-area adequacy does not establish transmission deliverability, operating security, or customer outage performance.

Example interpretation
“The portfolio’s mean LOLE is below 0.1 shortfall-day/year, but the 95% interval crosses the criterion. Risk is concentrated in September evenings and in two weather years. The result is therefore sampling-uncertain and weather-sensitive; more outage draws and targeted sensitivity cases are warranted before concluding that the system meets the reference.”

That statement is more decision-useful than “the model passed.”

10 · Professional habits

Questions an industry analyst should ask

01
What exactly is the metric and event period?

Days, hours, contiguous events, MWh, ppm, or affected years are not interchangeable.

02
What uncertainties are sampled?

Weather, load forecast, outages, hydro, imports, ambient derates, fuel, and common-mode events may be treated differently.

03
How many independent weather observations exist?

Repeated outage draws improve conditional estimates but do not expand the weather record.

04
Is capacity deliverable?

A single-zone model assumes away transmission constraints that may prevent resources from reaching load.

05
How are storage and demand response dispatched?

Starting SOC, energy budgets, charging, foresight, and event duration can materially change capacity value.

06
Is the conclusion stable?

Check confidence intervals, convergence, weather-year dispersion, sensitivities, and alternative risk metrics.

07
What benchmark validates the model?

Study-grade work should reconcile documented test systems or published cases before informing procurement.

Primary reading

Continue with industry sources