Distribution Network Assessment using a Causal-Guided Diffusion Model of Synthetic Residential Load Profiles

Motivated by the need for decarbonization, power systems are increasingly dependent on advanced models to build reliable power distribution networks. As these distribution networks become increasingly data-driven, customer load profiles represent valuable resources for operational characterization and decision-making. Nevertheless, these datasets are impacted by privacy, economic and quality barriers, challenging their accessibility and re-use. Against this backdrop, the new range of high performing generative models have shown increasing ability to produce realistically synthetic data. Despite their potential, these models are challenged by their inherent complexity and opacity, particularly when processing high-dimensional and stochastic data. Consequently, this absence of transparency offers limited operational value when models are developed in a standalone manner, as they cannot provide the actionable guidance necessary for operators to make informed decisions. To address these challenges, this paper introduces a causal-guided generative model coupled with a power flow simulation for disentangled guidance of realistic synthetic smart meter data to support distribution network analysis and planning. The proposed approach combines a causal inference module with a conditional diffusion model to generate load profiles that are statistically accurate and causally consistent in both social practice and electrically coherent terms. This framework offers a better understood means of reliably augmenting limited smart meter datasets, with potential to support demand-side flexibility assessment, reinforcement studies and the evaluation of policy interventions across the power distribution system.

Keywords: Causal inference, LV networks, interpretability, Generative AI, Decision-making, reliability.