Overview
A forecasting model for the energy behaviour of prosumers, customers who both
consume and generate electricity. Built at Wageningen in January 2024 on
large-scale smart-meter data.
The problem
Prosumers make grid balancing harder. A household with solar panels is a load for
part of the day and a supplier for another part, and the switch depends on
weather, season and behaviour at once. Imbalance costs money, so a forecast that
is wrong in a predictable direction is worse than useless.
What I did
- Engineered time-series and weather features over large-scale smart-meter data.
- Built a gradient-boosting pipeline using LightGBM and CatBoost.
- Set up cross-validation with early stopping and tuned hyperparameters against
mean absolute error.
Tech stack