DiffCharge: Generating EV Charging Scenarios via a Denoising Diffusion Model

TL;DR

The paper tackles the challenge of modeling uncertain EV charging loads that stress power grids and proposes DiffCharge, a denoising diffusion probabilistic model that generates both battery-level charging curves $\mathbf{r}\in\mathbb{R}^{L_r}$ and station-level daily loads $\mathbf{s}\in\mathbb{R}^{L_s}$, using a hybrid LSTM-Transformer denoising network with conditional generation. It trains with a quadratic noise schedule and ELBO-based objectives to learn step-wise noise $\bm{\epsilon}$, enabling high-fidelity, diverse time-series samples that preserve key charging dynamics such as the bulk and absorption stages and station-specific patterns. The approach demonstrates superior qualitative and quantitative performance over GMM, VAEGAN, and TimeGAN baselines, including realistic duration distributions and absorption-tail features, and it can condition on station type to produce distinct load profiles. A practical day-ahead market bidding case shows that DiffCharge-generated scenarios yield bidding plans close to real data, illustrating meaningful utility for grid operations and EV charging management under uncertainty. Overall, DiffCharge advances realistic EV charging scenario generation, providing a scalable tool for uncertainty-informed planning and operation in power systems with high EV penetration.

Abstract

Recent proliferation of electric vehicle (EV) charging events has brought prominent stress over power grid operation. Due to the stochastic and volatile EV charging behaviors, the induced charging loads are extremely uncertain, posing modeling and control challenges for grid operators and charging management. Generating EV charging scenarios would aid via synthesizing a myriad of realistic charging scenarios. To this end, we propose a novel denoising Diffusion-based Charging scenario generation model DiffCharge, which is capable of generating a broad variety of realistic EV charging profiles with distinctive temporal properties. It is able to progressively convert the simply known Gaussian noise to genuine charging time-series data, by learning a parameterized reversal of a forward diffusion process. Besides, we leverage the multi-head self-attention and prior conditions to capture the temporal correlations and unique information associated with EV or charging station types in real charging profiles. Moreover, We demonstrate the superiority of DiffCharge on extensive real-world charging datasets, as well as the efficacy on EV integration in power distribution grids.

Figures (15)

• Figure 1: Two-stage battery charging curves with distinct temporal patterns.
• Figure 2: Charging load profiles in different stations. (a) Workplace. (b) Campus.
• Figure 3: Overview of the proposed DiffCharge framework based on the denoising diffusion model for EV charging scenario generation.
• Figure 4: The architecture diagram of the denoising network $\bm{\epsilon}_{\theta}(\cdot)$. For a more clear display, we set the length of time steps $L$ in $\mathbf{x}_{t}$ as 4 in this graph.
• Figure 5: Weights $\sqrt{\alpha_{t}}, \sqrt{1-\alpha_{t}}$ associated with mean and variance term presented in \ref{['eq:4']} under the quadratic noise schedule.