Abstract:Generating synthetic data has become a popular alternative solution to deal with the difficulties in accessing and sharing field measurement data in power systems. However, to make the generation results controllable, existing methods (e.g. Conditional Generative Adversarial Nets, cGAN) require labeled dataset to train the model, which is demanding in practice because many field measurement data lacks descriptive labels. In this paper, we introduce the Information Maximizing Generative Adversarial Nets (infoGAN) to achieve interpretable feature extraction and controllable synthetic data generation based on the unlabeled electrical time series dataset. Features with clear physical meanings can be automatically extracted by maximizing the mutual information between the input latent code and the classifier output of infoGAN. Then the extracted features are used to control the generation results similar to a vanilla cGAN framework. Case study is based on the time series datasets of power load and renewable energy output. Results demonstrate that infoGAN can extract both discrete and continuous features with clear physical meanings, as well as generating realistic synthetic time series that satisfy given features.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: in the power system, due to considerations of energy security and user privacy, it has become very difficult to obtain and share on - site measurement data. To meet this challenge, generating synthetic data has become an alternative. However, existing generation methods (such as Conditional Generative Adversarial Networks, cGAN) require labeled datasets to train models, which is very difficult in practical applications because many on - site measurement data lack descriptive labels. To solve this problem, this paper introduces the Information - Maximizing Generative Adversarial Network (infoGAN) to achieve interpretable feature extraction and controllable synthetic data generation based on unlabeled power time - series datasets. By maximizing the mutual information between the input latent code and the output of the infoGAN classifier, features with clear physical meanings can be automatically extracted. Then these extracted features are used to control the generation results, similar to the traditional cGAN framework. ### Main problem summary: 1. **Difficulty in obtaining and sharing on - site measurement data**: Due to energy security and user privacy issues, it is very difficult to obtain on - site measurement data in the power system. 2. **Existing methods rely on labeled data**: Existing generation methods such as cGAN require labeled datasets for training, and much data in reality lacks labels and is difficult to obtain. 3. **Improving the controllability and interpretability of generated data**: A method is needed to generate synthetic data with clear physical meanings and controllability without relying on labeled data. ### Solutions: - **Introducing infoGAN**: Use the Information - Maximizing Generative Adversarial Network (infoGAN) to achieve unsupervised learning, extract interpretable features from unlabeled power time - series datasets, and generate controllable synthetic data. - **Combining feature extraction and generation**: By maximizing the mutual information between the latent code and the generated data, extract features with physical meanings, thereby achieving control over the generation results. Through these methods, researchers hope to generate realistic synthetic time - series data while maintaining data diversity and controllability, thus promoting the development of data - driven technologies in the power system.

Unsupervised and Interpretable Synthesizing for Electrical Time Series Based on Information Maximizing Generative Adversarial Nets

Synthetic Active Distribution System Generation Via Unbalanced Graph Generative Adversarial Network.

Generating Synthetic Mixed-Type Longitudinal Electronic Health Records for Artificial Intelligent Applications

Data-driven Scenario Generation of Renewable Energy Production Based on Controllable Generative Adversarial Networks with Interpretability

Synthetic Time-Series Load Data via Conditional Generative Adversarial Networks

TTS-GAN: A Transformer-based Time-Series Generative Adversarial Network

Conditional-TimeGAN for Realistic and High-Quality Appliance Trajectories Generation and Data Augmentation in Nonintrusive Load Monitoring

Synthetic Dynamic PMU Data Generation: A Generative Adversarial Network Approach

InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets

zGAN: An Outlier-focused Generative Adversarial Network For Realistic Synthetic Data Generation

Time-series Transformer Generative Adversarial Networks

Variational Autoencoder Generative Adversarial Network for Synthetic Data Generation in Smart Home

Synthetic Data Generation for Residential Load Patterns via Recurrent GAN and Ensemble Method

Learning to Generate Time Series Conditioned Graphs with Generative Adversarial Nets

GAT-GAN : A Graph-Attention-based Time-Series Generative Adversarial Network

Generative Adversarial Networks Applied to Synthetic Financial Scenarios Generation

STAN: Synthetic Network Traffic Generation with Generative Neural Models

Forecasting Renewable Energy Generation Scenarios Based on Multi-Agent Diverse GANs

Generation of Realistic Synthetic Financial Time-series

Are Synthetic Time-series Data Really not as Good as Real Data?

Can GANs Learn the Stylized Facts of Financial Time Series?