Fleet reliability

Failure

An equipment fault on one asset that needed someone on site or a part. Only failures count in failure rates, MTBF and MTTR.

Trip

An equipment trip that was reset remotely, with nobody attending. Trips have their own rate, worked out the same way as the failure rate. A grid-disturbance trip counts as external only if the dip was outside the grid code ride-through envelope; otherwise it is a plant failure or trip.

Failures to start

Engine start failures and gas valve proving failures are failures on demand. They are counted per start, from each engine's start count, with an exact 90% interval, and kept out of the running-hour failure and trip rates. Their downtime and cost stay in the totals.

Left out

Planned maintenance and service overruns, external causes (grid, curtailment, the DNO, gas supply, weather stows), and setting or maintenance errors with no equipment fault (protection settings left wrong, set points not released, strings left disconnected after maintenance). They are on the loss waterfall; the setting and maintenance errors are counted by type under At a glance, because in IEC 60050-192 terms they are systematic failures, not random ones. Equipment classes with no failures stay in, with 0 and their upper bound. The asset register also holds items that are not equipment here: .

Periods

Failures count in the period they start; downtime, energy and revenue are split at period ends, as on the loss waterfall. Work orders count with their failure.

Exposure

Unit-years in service: the sum of each unit's hours in service in 2025, divided by 8,760. Every unit in the register counts, including units that never failed, because they are what makes a rate. For a quarter or a month, each unit's hours in service are spread evenly over the part of the year it was in service.

Failure rate

Failures divided by unit-years. The 90% interval is the two-sided chi-squared interval for time-truncated data, assuming a constant rate: lower bound χ²(0.05, 2r) / 2T and upper bound χ²(0.95, 2r + 2) / 2T, for r failures in T unit-years. The page computes the chi-squared quantiles itself, by inverting the regularised incomplete gamma function. With few failures the interval is wide, so models are never ranked on the rate alone.

Comparing models

Only within a class: a central inverter and a string inverter are not comparable per unit. Each model is set against the rest of its class in view, as a ratio of rates with an exact 90% interval from the conditional binomial test. A model is called clearly above its peers only when the lower end of that interval is above 1. Where unit sizes in a class differ by more than half again (a string input of five strings against one of a single string, say), it must also be above 1 per MW of rating, because a bigger unit has more parts to fail. Where each model sits at its own site, the comparison mixes the model with the site's climate, installation and O&M, and failures with no root cause recorded weaken it further. With many models tested, one can cross the line by chance, so a finding is a reason to look, not a verdict.

Across classes

Energy lost to failures and trips, energy-based unavailability where the data has each unit's expected energy, and time-based unavailability are the measures that can be set side by side across classes. A per-unit rate cannot. The headline unit-hours down and mean time to restore pool every class, so an hour of a 12 kW string counts the same as an hour of a 3 MW turbine; the energy-weighted figure corrects for that.

MTBF

Mean time between failures: unit-hours in service divided by failures. For calendar-hour units it does not take the downtime out, because downtime is a small share of hours in service ( for the worst model in this data). For the engines, hours in service are the scheduled run windows, which include downtime. One failure per 70 unit-years is a rate across a fleet, not the life of one unit. A model with no failures has no MTBF; its rate interval shows how long it might be.

MTTR

Mean time to restoration, in IEC 60050-192 terms: the mean time to restoration of failures, from the alarm to the asset back in service. It is split into waiting for attendance, and waiting for spares or access (a part, a crane, a weather window), which are logistic delay with any administrative delay inside them, and repair, which is the active corrective maintenance time. For each failure the three add up to its hours out of service. The energy-weighted figure weights each failure by the MWh it lost. Time before the alarm, when a fault was not yet detected, is not measured. Failures still open at the end of the year are left out. It is not the mean repair time, which would count only the hands-on repair.

Unavailability

Time-based: unit-hours down for failures and trips divided by hours in service, with overlapping events on one unit counted once. For the engines the base is the scheduled run windows. Energy-based, where the data has each unit's expected energy: energy lost to failures and trips over expected energy, for the full year only.

Cost of downtime

Work order cost (labour, parts, access such as cranes, and other costs such as call-out charges) plus lost revenue, for each event. Lost revenue uses the same valuation as the loss waterfall's £ view: the site tariff, or the CHP's own margin. Who bore each cost comes from the work order: labour, access and other costs sit with the party named for the work order, and parts with the party named for the parts. Owner's cost counts only what the owner bore: those work order items where the owner is named, plus all lost revenue, before any availability guarantee payment. List prices count every work order in full, including costs inside OEM service fees and warranties.

Disputed

The party responsible under the contract does not accept the event or how it is classed, for example as a grid event or an owner stop. The event detail gives the reason, as on the loss waterfall. Disputed events stay in every figure here.

Did the repair work

For each failure that started in the period and whose work order has closed, the asset's time-based availability over the 30 days before the failure and the 30 days after the work order closed. Failures, trips and planned work count as down; time lost to external causes on the asset or at the site, such as a grid outage, is taken out of the window. A repeat is another failure of the same asset in the same subsystem that starts after the asset was restored and before the window after closure ends. The repeat share counts only repairs whose window ended inside the data; its 90% interval is exact (Clopper and Pearson). A repair followed only by a trip in the same subsystem is counted on its own. To compare repair methods, a repeat is the same failure mode on the same asset, so a pitch fault on another blade does not count against an accumulator replacement. Methods with at least two repairs each are compared with a one-sided Fisher's exact test; a method is suggested only below p = 0.1, and called proven only below p = 0.05.

Bad actors

Units ranked by failures, then downtime. At a constant rate, each unit's failures would follow a Poisson distribution with a mean set by its share of the hours in service, so the page counts the units that failed more than once and sets that against the number a constant rate would give. Many more than that points at those units, not only the model.

Same events as the loss waterfall. Every event here is an event on the generators loss waterfall, with the same id, start, end, hours and MWh. An event the loss waterfall splits at a month end is one record here, with its month parts. Some events are booked to a different unit from the one the waterfall names: a communications loss to the logger or network switch rather than the inverter that dropped out, and a DC insulation fault to the string rather than the inverter that tripped. So cross-check by event id, not by asset.

Labour rates.

Part and access prices. Each work order's parts are priced from this list; access (cranes, excavation, rope access) is listed with the parts.

Sample data. The seven sites, their names, events, work orders, costs and every number on this page are fictional, generated to show how the report behaves over the full year 2025. Equipment models are named to make the sample realistic; the faults and figures attached to them are invented and say nothing about the real products.