Reliability and resilience in electrical network planning

Most networks rank investments on expected value. For reliability that works. For resilience it quietly fails, and the failure is systematic: you underfund the tail without ever deciding to.

Reliability, resilience, DER (Distributed Energy Resources) readiness, digital modernization and growth all compete for the same finite capital. You know this. What’s less obvious is that the arithmetic most organizations use to arbitrate between them is not neutral. It has a bias, the bias runs in one direction, and it is large enough to invert a capital plan.

Expected value works for reliability. It breaks for resilience.

Turning risk into a convertible currency is the right instinct, and we set out the mechanics of it in the first article in this series. You reduce each risk to the same expected-value form: probability of a bad outcome times its monetized consequence. A dollar of reliability risk bought down becomes directly comparable to a dollar of resilience or growth risk. Same units, one ranking, no politics.

Then you try it, and the distributions don’t match.

Reliability risk is frequent and well-behaved. You have hundreds of failures a year and decades of records, so the mean is a genuinely good description of what next year will cost you. Resilience risk is the opposite: low probability, severe consequence, fat-tailed. A one-in-fifty-year event has an expected annual cost that looks modest and a realized cost that can exceed a decade of reliability spending. The mean is not a summary of that risk. It is an average of a scenario that mostly doesn’t happen and occasionally dominates everything.

So a serious framework adds a tail measure. But here the standard fix has a trap in it, and it’s worth walking through the numbers.

A worked example: how the ranking inverts

TTwo programs compete for the same capital on the same network. The parameters below are illustrative, chosen to be round and checkable rather than drawn from any client engagement. Run your own numbers and the pattern holds.

The reliability program addresses an asset class with roughly 200 failures a year at an average monetized consequence of $50,000. It costs $3 million a year and halves the failure rate. The resilience program hardens the network against a severe ice storm carrying a 2% annual probability and a $400 million consequence. It costs $12 million a year on an annualized basis and cuts the consequence to $100 million.

Illustrative parameters. Single event class, single year, no correlation between programs.

Read the second row carefully, because it is the part that catches people. At a 95% confidence level, value at risk reports the resilience exposure as zero. Not understated. Zero. Ninety-eight percent of years contain no storm, so the 95th percentile year is a quiet one. An organization that adopts VaR specifically to correct the expected-value problem can still make its single largest exposure disappear from the ranking.

The fix is a conditional tail measure. Tail value at risk asks a different question: given that we are in the worst 5% of years, what is the average loss? For the resilience program that averages the storm years against the quiet ones inside the tail and returns $160 million. For the reliability program, where the law of large numbers keeps the distribution tight, it returns about $11.5 million.

Now the part that decides the plan, because gross exposure is not a capital decision. Nobody funds the larger number. They fund the better buy, which means risk reduced per dollar spent. On expected value the reliability program returns 1.7 against the resilience program’s 0.5, so reliability wins the capital and it is not close. On tail value at risk the resilience program returns 10.0 against reliability’s 1.8, so resilience wins by more than five to one. Same portfolio, same year, same data, same costs. The choice of risk measure, not the choice of investment, decided the plan.

One honest caveat, because a careful reader will find it. The $160 million figure depends on where the confidence level is set: $80 million at 90%, $400 million at 98%. That is a factor of five driven by a threshold choice rather than by anything about the network. What survives the threshold is the ranking. Across that whole range the resilience program stays the better buy on a tail measure and the worse buy on expected value. The magnitude is a modelling choice that has to be stated. The inversion is not.

Two further framing choices deserve to be explicit. The table covers a single year, which flatters the reliability program, since over a twenty-year horizon the probability of at least one severe storm is 33%, not 2%. And the programs are treated as uncorrelated, which is generous, because a storm year is also a year of elevated equipment failure.

This is why the measure, the confidence level, the horizon and the correlation assumptions all have to be explicit, documented decisions rather than defaults inherited from whichever spreadsheet came first.

The frontier is built from whichever measure you chose

From here the portfolio-theory machinery does what you’d expect. For a given level of risk, which allocation produces the most value? Plot the candidate portfolios and they trace an efficient frontier: the allocations where you can’t buy more value without taking more risk. An explicit value framework, the organization’s objectives translated into weighted and quantified measures at present value, ranks competing investments on one scale of merit such as NPV or benefit-cost ratio. Then you pick the point on the frontier matching your risk appetite.

But the frontier inherits the bias of its inputs. Build it on expected value and it is a real frontier of the wrong problem. That is the uncomfortable implication of the example above: an optimization can be technically flawless and still be optimizing a portfolio you would not have chosen had the tail been visible.

The frontier also isn’t a single deterministic line, which is how it is almost always drawn. Monte Carlo simulation across thousands of allocation scenarios, spanning different deterioration trajectories, demand forecasts and climate realizations, returns a probabilistic envelope. That distinction has a decision attached to it. The portfolio sitting on the frontier in your base case may not hold its position when the future moves. A slightly less efficient allocation that stays on the frontier across most scenarios is usually the better plan, and you cannot see that trade-off on a deterministic curve.

The lever nobody prices: timing

Money has a cost, so deferring an investment that doesn’t need to happen yet carries real present-value benefit. The question is never only what to build but when. An intervention that’s optimal for one asset in isolation may lose out once it competes for funds with everything else. Two plans with identical projects can deliver materially different value through sequencing alone: one defers cost and releases capacity at the right moments, the other doesn’t. Optimizing the calendar, not just the list, is often the cheapest improvement available. It costs analysis, not steel.

In Australia, CitiPower did this across 103 primary transformers. Rather than defend a replacement date, it calculated for each unit the year in which discounted investment cost plus accumulated risk was lowest, then scheduled to that year. Stakeholders pushing to delay replacement got an answer in numbers instead of an opinion.

Three questions to test your own process

None of this is abstract. Each of these has a yes or no answer, and a no is a finding.

1. Can you state resilience risk and reliability risk in the same units today?

Not describe both well. State both as monetized risk, on the same scale, in a document someone outside your team could audit.

2. Does your process choose the investment year, or do you hand it the year?

If the year arrives as an input from the asset team, timing is not being optimized. It is being assumed, and the present-value benefit of deferral is invisible to the ranking.

3. Is your portfolio on the frontier across scenarios, or only in the base case?

If you have never tested where your allocation lands under a harsher deterioration curve or a wetter climate scenario, you know it is efficient under one future and nothing about the others.

A no on the first question means the tail problem in the example above is currently invisible to your capital plan. A no on the second means you are leaving present value on the table. A no on the third means your plan is optimized for a forecast rather than a range.

Why this is becoming unavoidable

Networks fund reliability, resilience, electrification, large industrial loads and digital modernization at once, with capital that hasn’t grown to match. ISO 55000, the international standard, frames asset management as balancing cost, risk and performance against organizational objectives. What has changed is that the balance now has to be quantified, and networks must justify the portfolio rather than each project in it. It is no longer enough to show each investment is individually sound. You have to show the mix is the best available use of finite capital, and that no other combination or sequence would deliver more value for the risk carried. The framing of the plan is under the same scrutiny, which is the subject of the third article in this series.

Regulators are already moving. In the United Kingdom, Ofgem has harmonized how the companies it regulates report risk across electricity distribution, electricity transmission and gas, using asset health and criticality to arrive at monetized risk comparable between companies. In Ontario, the Energy Board requires distributors to file asset management plans setting out their planning process and lifecycle optimization practices. Québec’s public infrastructure framework requires public bodies to quantify their inventory, asset condition, maintenance needs, deferred maintenance deficit and replacement value. Different jurisdictions, same demand: show the numbers, not the narrative.

Which brings the tail problem into regulatory territory. Once a regulator can set your risk figures beside your peers’, a framework that reports your largest exposure as zero is not a modelling preference. It is an exposure of its own. The same bar now reaches municipal and public asset owners from a different direction.

How Direxyon does this

A value-based multi-criteria framework, using methods such as TOPSIS, ELECTRE or MACBETH, reduces reliability, resilience and growth risk to the same risk-adjusted units. Fat-tailed resilience risk is carried on a conditional tail measure rather than the mean or a naive percentile, so the exposure that would otherwise vanish stays in the ranking. The method optimizes investment timing alongside the mix, and Monte Carlo simulation returns the frontier as a probabilistic envelope rather than a single curve, so robustness is visible next to efficiency. The output is one auditable ranking, with every trade-off documented, in a form that holds up in front of a regulator.

Frequently Asked Questions