Forecasting against persistence in data-poor grids.

The baseline nobody can skip

Persistence is the forecast that says tomorrow will look like today: the same solar yield, the same load shape, the same site behaviour. It requires no model, no training data and no expertise, and it is embarrassingly hard to beat, because in stable climates and stable communities, tomorrow usually does look like today.

That makes persistence the honest floor for any forecasting claim. A model that cannot outperform persistence on real data from the site it claims to serve is not a capability. It is a demo. We hold our own forecasting work to that bar, and we state the comparison alongside the forecast rather than in a footnote, because an operator deciding whether to trust a number deserves to know what it beat.

Why data-poor grids change the problem

Forecasting literature mostly assumes the luxuries of a data-rich grid: years of clean telemetry, dense weather stations, reliable timestamps. A rural mini-grid offers none of these. Histories are short, because the site was energised eighteen months ago. Telemetry has gaps, because the gateway shares the site's connectivity problems. Weather ground truth may be a station a hundred kilometres away describing a different microclimate.

Three consequences follow. First, model complexity is mostly wasted; with short and gappy histories, a heavily parameterised model learns the gaps, not the site. Simple models with strong physical priors, such as clear-sky irradiance shaped by satellite-derived cloud cover, tend to travel better than deep architectures fed thin data. Second, satellite and reanalysis sources matter disproportionately, because they are the only inputs whose availability does not depend on hardware at the site. Third, the evaluation itself must respect the gaps: scoring a forecast only on the days when telemetry happened to arrive quietly biases the result toward the site's good days.

Forecasts carry their own error

A forecast published as a bare number invites exactly the misuse it will eventually get: someone will schedule diesel, or promise a customer, or size a battery against it. Our practice is that a forecast ships with its recent error against persistence, stated in the same units the operator thinks in. When the model's edge over persistence narrows, the system says so. When the inputs that give the model its edge go missing, the forecast degrades explicitly toward persistence rather than pretending otherwise, and the degradation is visible on the screen.

The uncomfortable conclusion

Some sites, some seasons and some horizons do not reward a model at all. Where persistence is genuinely the best available forecast, the honest system presents persistence, labelled as such, and spends its sophistication elsewhere: on detecting when conditions are departing from yesterday, which is the one thing persistence is structurally unable to do. Knowing when the simple answer is the right answer is itself a forecasting capability, and it is the one most demos skip.