Multi-Dimensional Time-Series Data Synthesis Based on Improved TimeGAN

Artificial intelligence models are increasingly used in industrial applications to assist with predictive tasks. However, the complexity of industrial environments poses challenges to data collection, transmission, and storage, often leading to data loss and insufficiency. Synthetic data generation has become a common solution to augment datasets. Existing methods can be broadly categorized into traditional statistical techniques and deep generative approaches. Statistical methods explicitly model observed data, proving advantageous for simple, small-scale synthesis tasks. In contrast, deep generative methods implicitly model data and are capable of generating more complex datasets. Nevertheless, the data synthesized by deep generative models can be uncontrollable, with issues such as mode collapse. This paper proposes a novel data synthesis method based on the deep generative model TimeGAN, aimed at synthesizing multi-dimensional wind power time-series data. TimeGAN combines the architectures of GAN and AE (Autoencoder). The proposed method introduces a parallel structure of TCN (Temporal Convolutional Networks) and Transformer as the backbone of the generator and embedder. Additionally, Wasserstein divergence is used in place of KL (Kullback-Leibler) divergence to calculate loss, effectively mitigating mode collapse during training. Comparative experiments with baseline algorithms, including TimeGAN and diffusion models, demonstrate the effectiveness of the proposed method. Furthermore, this work provides new insights into predictive tasks where data collection is difficult or incomplete.