Energy Nowcasting & Forecasting for LEAP

Problem

LEAP streamed electricity consumption, gas consumption and solar generation data every 15 minutes across the platform’s data space, the same live data space Cooee’s chatbot would later sit on top of. That stream told you what had already happened. It didn’t tell you what was about to happen, which is the question that actually matters for managing load, catching anomalies early or planning around generation.

Approach

• Nowcasting: short-horizon models estimating the current or near-immediate state from the latest readings, for situations where waiting for the next 15-minute interval to confirm a trend is already too late.
• Forecasting: longer-horizon models over the historical series, predicting usage, solar generation and emissions ahead of time rather than just reporting them after the fact.
• Model selection per signal: gradient-boosted trees (XGBoost, LightGBM) where the relationship between features and target was more tabular than sequential, and neural sequence models, including LSTMs, where the time-series structure itself carried most of the signal.

Treating nowcasting and forecasting as distinct problems, rather than one model asked to do both, mattered because they fail differently: a nowcast that’s slightly off is a rounding error, a forecast that’s slightly off compounds over the horizon.

Result

Working nowcasting and forecasting models for energy usage, solar generation and emissions, built on the platform’s live 15-minute smart meter stream, feeding the same data space that LEAP’s public and internal tooling draws on.

What I’d do differently

I’d formalise a rolling backtest from the start rather than evaluating primarily on a fixed holdout. Energy consumption patterns shift with seasons, occupancy and infrastructure changes, and a model validated once on a static split can look fine in testing while quietly drifting in production.