Inferring and embedding behavioral patterns inside the forecasting model—rather than using them only for grouping—improves electricity load predictions for diverse households, especially when you have limited historical data.
This paper improves residential electricity demand forecasting by embedding behavioral patterns directly into a Neural Process model. Instead of just grouping similar households, the model learns discrete behavioral structures (like weekday vs. weekend routines) and uses them to condition predictions, while also capturing uncertainty across different household types.