LoaDiff: Conditional Generation of Electricity Consumption Time Series for Energy Analytics

TL;DR

LoaDiff is a diffusion model that generates realistic, year-long, sub-hourly electricity load curves conditioned on household attributes and environmental factors like temperature. It produces diverse and high-fidelity synthetic data while maintaining utility for downstream tasks such as load forecasting and appliance detection.

Why It Matters

Synthetic smart-meter data is essential for advancing energy analytics without violating privacy regulations. LoaDiff enables researchers and utilities to generate realistic consumption patterns for training models and testing demand-response strategies, accelerating innovation in energy systems.

Key Findings

• LoaDiff generates highly realistic and diverse electricity load profiles at sub-hourly resolution over full years.
• The model maintains low memorization risk, avoiding overfitting to training data while preserving useful statistical patterns for downstream applications.
• Conditional generation based on temperature and household characteristics produces coherent and plausible responses, enabling scenario-based energy planning.