Generative modeling for long-term load forecasting with customer energy asset adoption trends

초록

This paper presents a data-driven approach for estimating future distribution grid load projection scenarios driven by increasing customer adoption of electric heating systems and customer-owned distributed generation, without requiring sub-metered data for individual load components. The proposed method integrates load disaggregation and generative load modeling into a unified framework. First, we disaggregate the net load into a base load, heat ing, ventilation and air-conditioning (HVAC) load, and photovoltaic (PV) generation, producing component-wise load profiles essential for future scenarios such as widespread adoption of heat-pumps and increased behind-the-meter generation penetration. We define an adoption proxy capturing trends and linking to component-wise load to model plausible future load projections. Building on recent advancements in diffusion models, we intro duce a conditional diffusion-based framework that jointly models daily profiles of net load, HVAC load, and PV generation as random variables. The proposed diffusion model enhances sampling efficiency through accelerated reverse diffusion and improves modeling capability using a diffusion transformer network, where condition ing inputs are reasoned through cross-attention and self-attention, improving information flow. This generative framework enables the prediction of load scenarios that capture the incremental impact of each component while accounting for uncertainties. The model, trained on aggregated advanced metering infrastructure data provided by the utility, demonstrates around 30% higher accuracy in generating future load scenarios compared to the leading generative and discriminative models.

키워드