Forecasting load across 2.1 million smart meters

Forecasting ran through a spreadsheet-shaped pipeline that one analyst maintained and, realistically, one analyst understood. Meter reads arrived as nightly CSV drops, so the day-ahead position was built on data up to twenty hours old. Error sat at 9.4% MAPE, and every point of that error gets bought back on the balancing market, usually at the worst hour of the day. That is how a data-freshness problem ends up on a P&L.

We replaced the drops with a streaming ingest that takes reads at fifteen-minute cadence and back-fills late arrivals without corrupting a sealed settlement window — meters drop offline and return with four hours of history behind them, and the old pipeline simply lost it. Forecasting moved to a per-substation ensemble: gradient boosting over weather, calendar and local load history, sitting on a shared regional temporal model. It retrains nightly and has to beat the incumbent on a held-out week before it is allowed to promote itself. Every forecast ships with its input attribution, because the first question from the trading desk is never what the number is — it is why it moved.

Day-ahead error went from 9.4% to 6.5% MAPE, and the intraday position now updates hourly rather than never. Solstice puts avoided balancing cost at roughly $4.6M a year; we would call that directionally right rather than precise. The pipeline has held its fifteen-minute freshness SLO unattended for eleven months, through two substation outages and one very cold February.