22 May 2025 in Energy Economics
9 April 2026 in SSRN Electronic Journal
The electrification of residential energy use through heat pumps, electric vehicles, and rooftop solar increases the variability and heterogeneity of electricity demand and generation. For system operators and electricity market participants, timely information on household technology adoption is essential but often unavailable. This paper examines whether smart meter data can be used to infer low-carbon technology (LCT) adoption under realistic data constraints. We assess whether short observation windows per household—one week or one month—are sufficient to predict detailed technology adoption groups. Using public, anonymized 15-minute smart meter data for 2,400 Flemish households, evenly distributed across eight adoption groups of combined LCTs, we apply a range of unsupervised and supervised machine-learning methods at different time aggregation levels. Results using yearly data show strong predictive performance, with macro-average F1 scores up to 0.75, reliably detecting complex combined adoption of electric vehicles, heat pumps and rooftop solar. We extend the analysis to a larger proprietary dataset of Flemish households, handlabelled using administrative records, where macro-average F1 increases to around 0.79. Crucially, accurate detection is also achievable for most classes with just a few random sequential months of data per household. Our publicly available model shows that complex LCT combinations can be efficiently detected from limited consumption data, balancing predictive performance, timeliness and computational complexity.
14 May 2026 in SSRN Electronic Journal
The electrification of residential energy use through heat pumps, electric vehicles, and rooftop solar increases the variability and heterogeneity of electricity demand and generation. For system operators and electricity market participants, timely information on household technology adoption is essential but often unavailable. This paper examines whether smart meter data can be used to infer low-carbon technology (LCT) adoption under realistic data constraints. We assess whether short observation windows per household—one week or one month—are sufficient to predict detailed technology adoption groups. Using public, anonymized 15-minute smart meter data for 2,400 Flemish households, evenly distributed across eight adoption groups of combined LCTs, we apply a range of unsupervised and supervised machine-learning methods at different time aggregation levels. Results using yearly data show strong predictive performance, with macro-average F1 scores up to 0.75, reliably detecting complex combined adoption of electric vehicles, heat pumps and rooftop solar. We extend the analysis to a larger proprietary dataset of Flemish households, hand-labelled using administrative records, where macro-average F1 increases to around 0.79. Crucially, accurate detection is also achievable for most classes with just a few random sequential months of data per household. Our publicly available model shows that complex LCT combinations can be efficiently detected from limited consumption data, balancing predictive performance, timeliness and computational complexity.
5 September 2026 in Sustainable Energy, Grids and Networks