An Improved Adversarial Convolutional Transformer Model for Time Series Forecasting

Time series forecasting, as an important research topic in time series analysis, plays a crucial role in many fields. Traditional Transformer model often suffers from low feature dimension extraction and poor generalization ability when performing time series forecasting tasks. To address these issues, this paper proposes an Improved Adversarial Convolutional Transformer (IACT) model for time series forecasting. First, multi-dimensional features are extracted using the multi-head attention mechanism, and then more features are captured by multi-dimensional convolutional subspace in different dimensions to improve the representation capabilities of features. Meanwhile, an adversarial training strategy is used to make the model adapt to a larger range of data variations, which makes the model have a stronger generalization ability. Finally, experiment results conducted on two real time series datasets show that the model can obtain excellent prediction effects and generalization performance. In addition, the proposed model outperforms benchmark models such as Transformer, Pyraformer, and Informer in terms of prediction accuracy.