One year of history is one winter

Twelve months of ten-minute data is a substantial dataset by most measures: roughly fifty-two thousand records per channel per turbine. It is also, from the point of view of seasonality, a sample size of one. Every winter behaves like the one winter observed, every summer like the one summer, and a model has no way to distinguish a seasonal pattern from a trend that happens to have run for a season.

This matters because the two most valuable things a model can report have opposite requirements. A performance deficit that is present across all conditions is detectable in a few months, because the comparison is against other machines under matched conditions and the weather cancels. A slow degradation trend is a statement about change over time, and distinguishing it from an annual cycle needs at least two instances of that cycle.

The failure this produces is specific and worth naming. A gearbox temperature that rises from October to February is either a bearing degrading or a machine running harder in the windiest months at lower ambient temperatures, and with one year of history those hypotheses fit equally well. Two winters separate them immediately, because the pattern either repeats or continues. This is the reason the recommendation is twelve months as a working minimum and twenty-four as the preferred depth, rather than the twelve months most vendors ask for.

The counter-argument is that fleets do not always have the data, and waiting a year to begin is worse than starting with what exists. That is right, and it is how the sequencing works in practice: comparative findings are available from the first assessment, and trend-based condition findings become confident as history accumulates. The mistake would be presenting the second kind at full confidence in the first month, which is a straightforward way to send a crew after a seasonal pattern.

There is a second reason to want more history, less obvious than seasonality. Component replacements and retrofits are steps in a series, and a longer record contains more of them, which means it contains more examples of what a step looks like when it is maintenance rather than degradation. Older periods are weighted lower in the model - a fleet's behavior in 2016 is only partially informative about 2026 - but they are not discarded, because the value of a step is in what it teaches about steps.

The limit worth conceding is that history has diminishing returns and a real cost in relevance. A machine that has had a gearbox replaced, a controller upgrade, and a blade retrofit is in an important sense a different machine from the one commissioned twelve years ago, and modeling it against its own decade-old behavior would be a mistake dressed as thoroughness. More data is better only when the underlying asset has been stable enough for the data to describe the same thing.

For an operator deciding what to export, the practical guidance is: send everything the historian holds, tell us the major maintenance dates, and expect comparative findings first and condition trends to firm up over the following quarters. An assessment that promises confident degradation trends from twelve months of history is promising something the data cannot support.