Retraction Note: This paper was retracted by the publisher on November 12, 2025. The following summary describes the methodology and claims presented in the original text for informational purposes.
Abstract
High-quality demand-side management requires an abundance of load profiles to support decision-making processes. However, customer energy consumption data often contains sensitive personal information, and service providers face significant challenges in accessing a substantial amount of energy consumption data. To generate a large volume of customer data without compromising privacy, this study introduces a data-driven approach integrating Information Maximizing Generative Adversarial Networks (InfoGAN) with Multivariate Kernel Density Estimation (MKDE) for the generation of load profiles. InfoGAN is firstly trained based on existing customer load profiles, with the Q network disentangling the load into feature variables and the generator producing realistic profiles. Subsequently, MKDE is utilized to assess the distribution of these features, enabling the generation of new profiles by sampling new feature variables. The proposed method circumvents the need for intricate sampling or modeling processes and generates realistic data that represents the inherent uncertainties and fluctuations characterizing customers’ electricity consumption. The generated data could be used as the substitution for real electricity consumption data, thereby facilitating further applications without compromising privacy concerns.
1. Introduction
As smart grids evolve, Advanced Metering Infrastructure (AMI) collects massive amounts of fine-grained power consumption data. While this data is essential for demand-side management and grid optimization, it presents significant privacy risks. High-resolution load profiles can inadvertently reveal a customer’s daily habits, lifestyle choices, and even the presence of specific appliances. Because of these risks, researchers and service providers often struggle to access large-scale, real-world datasets. Traditional privacy-preserving methods, such as adding random noise or simple anonymization, often degrade the data's utility for complex modeling. This study aims to bridge this gap by generating "synthetic twins" of load data that maintain the statistical characteristics of the original data without containing actual personal information.
2. Methodology: The InfoGAN-MKDE Framework
The authors introduce a two-stage data-driven approach combining deep generative modeling with statistical density estimation.
2.1 InfoGAN and Feature Disentanglement
Unlike a standard Generative Adversarial Network (GAN), which maps random noise to data, InfoGAN includes an additional "Q network." This network is designed to "disentangle" the latent space, separating the input into incompressible noise and structured feature variables ($c$). These variables represent meaningful characteristics of the load profile, such as peak timing or consumption intensity. By identifying which latent variables control specific parts of the load curve, the model gains a level of interpretability that standard GANs lack.
2.2 Multivariate Kernel Density Estimation (MKDE)
Once the InfoGAN is trained, the authors use MKDE to model the joint probability distribution of the feature variables extracted by the Q network. Simply sampling from a standard Gaussian distribution often fails to capture the complex correlations between features. Using MKDE as a post-processing step allows the model to respect the actual statistical structure of the feature space. By sampling new feature variables from this MKDE-derived distribution and feeding them into the InfoGAN generator, the system can produce an infinite variety of new, realistic load profiles.
3. Experimental Validation
3.1 Datasets and Preprocessing
The study validated the model using a publicly available dataset of Small and Medium Enterprises (SME) in Ireland. The scope included data from 319 customers collected every 30 minutes over 1.5 years. The authors utilized 90,155 daily load profiles, normalized via min-max scaling. To test the model across different behaviors, they used K-means to group the data into three clusters: office-like (noon peaks), restaurant-like (evening peaks), and night-active enterprises.
3.2 Key Results
- Superior Similarity: The proposed method was compared against standard GANs and Variational Autoencoders (VAEs). Using the Maximum Mean Discrepancy (MMD) metric, the InfoGAN+MKDE approach consistently achieved lower scores, indicating a much closer match to real-world data distributions.
- Visual Fidelity: The generated profiles successfully captured the "spikiness" and inherent uncertainty of real electricity demand, which traditional mathematical models often smooth out.
- Small Sample Efficiency: The model demonstrated the ability to generate high-quality synthetic data even when provided with as few as 10 reference samples from a new customer type, making it flexible for niche applications.
4. Conclusion and Limitations
This hybrid framework offers a robust solution for generating privacy-preserving synthetic load profiles. However, the authors note that MKDE faces challenges with the "curse of dimensionality"; as the number of feature variables increases, the computational cost grows exponentially. Additionally, while the Q network disentangles features automatically, optimal selection of the number of feature variables remains an area for future exploration. Given the paper's retraction, readers should approach specific quantitative claims with caution, as retractions in this field often stem from issues with data integrity or methodological errors discovered post-publication.
Summary updated