MTMDC-GAN: Self-Attention Driven Multi-Scale Temporal Synthesis with Multi-Domain Analysis and Contrastive Learning

The synthesis of high-quality multivariate time series (MTS) is critical for enhancing the performance of predictive models and data-driven decision-making. This paper introduces a novel approach to MTS generation, integrating multi-scale data fusion with self-attention mechanisms to capture complex interdependencies across temporal scales, which to some extent aids the generator in producing coherent MTS, improving the representational capacity of the data. The methodology extends beyond conventional spatial feature extraction by incorporating Fourier Transformation analysis in the frequency domain and Contrastive Learning in the time domain. This multi-domain approach uncovers deeper insights into the underlying patterns and cyclic behaviors of time series data. Extensive experimentation across different datasets and configurations demonstrate superiority over existing techniques, achieving greater accuracy and realism in synthesized MTS.