Imputing Missing Values in Energy Consumption Time-Series

Sammanfattning

Missing data in smart-meter energy time-series disrupts downstream applications. This thesis frames missing data reconstruction as a multi-step imputation problem, employing a recursive framework to evaluate a globally trained XGBoost model against statistical baselines to impute gaps of up to 168 hours across three Swedish cities. Through a constructive ablation study, the impact of integrating temporal anchors, individualized household context, and exogenous variables like temperature and electricity spot prices was assessed. The results reveal a distinct metrics paradox: while the machine learning model provided the most robust general solution by minimizing relative percentage errors over week-long gaps, its tendency to generalize caused it to under-predict extreme physical peaks, an area where Seasonal Naive methods remained superior. Furthermore, integrating exogenous data did not guarantee improvement. Spot prices introduced noise, and temperature only enhanced accuracy when strictly paired with specific household identifiers. Ultimately, achieving high predictive accuracy required the model to memorize specific meter routines rather than learn universal energy behaviors.