Data Science
AutoTS Load Forecasting Pipeline
Automated load forecasting with AutoTS: genetic model search, rolling-window evaluation across a full year and template-based persistence of the best models.
What it’s about
A pipeline for forecasting the energy consumption of buildings and households (grid import in kW). It uses AutoTS, which searches across dozens of model classes with a genetic algorithm to find the best configuration. The sister project to my Darts pipeline: same task, different framework, different degree of automation.
Architecture
- Rolling-window forecasting with configurable horizon and stride (default: 6 hours): trained on January, evaluated on rolling forecasts from February through December
- Weekly retraining: AutoTS’ genetic search runs regularly, the best models are persisted via the template system (
export_template/import_template) and reused on the next run - Probabilistic forecasts: upper/lower bounds via a 95% prediction interval
- Six metrics per run: RMSE, NRMSE, MAPE, Pearson r, interval coverage and interval width
- SVG plots: actual vs. forecast with confidence band plus an error-difference plot, generated automatically per dataset
What I learned
AutoML for forecasting takes model selection off your plate, but not validation design. Horizon, stride and retraining frequency shape the result more than the model class does. Also: prediction interval coverage is an underrated metric. A model whose 95% interval only covers 70% of reality is more dangerous for energy management than one with a slightly higher RMSE.