A mid-sized manufacturer in Puducherry had done the sensible thing: put solar on the roof. But solar you can’t predict is solar you can’t plan around. Should tomorrow’s heavy, flexible loads run at midday under the array, or shift to avoid an expensive tariff window? Without a forecast, that decision was a daily coin-flip — and the answer was worth real money either way.

The obvious fix — hire a data science team, stand up a cloud pipeline — was wildly out of proportion for an MSME. They needed the forecast, not the org chart.

What we deployed

Vastvic Solar ran on the manufacturer’s own premises, reading their existing meter feed. From that history plus weather inputs, it produced day-ahead generation forecasts as calibrated P10/P50/P90 bands at 15-minute resolution — the same physics-informed quantile approach used for utility-scale plants, sized to a single rooftop.

From forecast to schedule

A forecast only matters if it changes a decision. The day-ahead generation curve fed a tariff-aware load schedule: flexible loads were nudged into the hours when on-site solar was likely to cover them, and away from the expensive grid windows. The P10 band mattered here — planning against a conservative estimate meant the schedule held up even on a cloudier-than-hoped day.

They didn’t get a data team. They got the one thing the data team would have been asked to build.

Why it fit

Everything ran on their infrastructure — no data leaving the site, no cloud dependency, no specialists to hire. The forecasts were calibrated, not optimistic, so operations could trust the ranges. And because the model is gated on beating a naive baseline before it ships, “the forecast is better than guessing” was a checkable claim, not a sales line.

Enterprise-grade forecasting doesn’t have to mean enterprise-grade overhead. A manufacturer got a day-ahead view of its own generation, a schedule that used it, and none of the machinery people assume you need to get there.

Key takeaways

  • Day-ahead P10/P50/P90 solar forecasts ran on the manufacturer’s own premises.
  • The forecast drove tariff-aware scheduling of flexible loads — planning against the conservative band.
  • No cloud dependency and no data team required — the model read the existing meter feed.
  • Calibration and a release gate made “better than guessing” a checkable property.