Long-Term, Multivariate Time Series Generation with the Capture of Intervariate Correlations and Variatewise Characteristics

This paper proposes a novel <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Time Series Generation (TSG)</i> model, the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Attended Variate-Conditioned GAN (AVC-GAN)</i>, for generating multivariate long-term time series data. Recently, generative approaches to TSG have been explored for applications such as privacy protection, anomaly detection, and time series classification/forecasting. Since existing TSG methods, particularly <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Generative Adversarial Networks (GANs)</i>-based methods, focus on modeling long-term time series and capturing intervariate correlations, they often struggle to adequately capture the characteristics of each variate, such as specific patterns of waveforms. To address these issues with TSG simultaneously, we propose a novel generative model, AVC-GAN. The key components of AVC-GAN are as follows: 1) a nonautoregressive architecture for modeling long-term time series, 2) variate conditioning for capturing variatewise characteristics, and 3) a multihead attention mechanism for capturing intervariate correlations. Our experiments on multiple benchmark datasets for <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Long-Term Time Series Forecasting (LTSF)</i> demonstrate thatAVC-GAN outperforms state-of-the-art GAN-based generative models across a range of evaluation metrics. In terms of distance-based metrics, calculated as the MSE between real and generated data, our method achieves improvements of 29.8% and 33.7% over state-of-the-art methods in assessing overall intervariate correlations and variatewise characteristics, respectively. Visualization also confirms the superior quality of the generated time series.