Generation of Household Electricity Demand Data Using Diffusion Models

We propose a novel household-level electricity demand data generation method based on Denoising Diffusion Probabilistic Models (DDPM), aiming to address the shortage of publicly available consumer-level datasets due to privacy concerns. By incorporating consumer-specific embeddings and multiple external factors, including temperature, day of the week, and holidays, our model generates high-fidelity demand patterns at 30-minute intervals. Experimental results on real-world data from 21,916 households demonstrate that the synthetic data reproduce both broad daily consumption trends and finer oscillations observed in the actual data. Furthermore, a forecasting model trained and validated on the generated data yields accuracy comparable to that obtained with the real dataset, suggesting that our approach can serve as a realistic and privacy-preserving platform for model validation. In addition to confirming the robustness of the method, we present how multiple variations of demand scenarios can be generated under the same conditions by altering initial random seeds. Future work will focus on adapting the model to broader temporal spans, exploring its applicability to other time-series domains, and investigating ways to compress and accelerate the diffusion model for large-scale adoption.