U-Net Inspired Transformer Architecture for Multivariate Time-Series Synthesis

This study introduces a Multiscale Dual-Attention U-Net (TS-MSDA U-Net) model, for long-term time-series synthesis. Enhancing the U-Net architecture with multiscale temporal feature extraction and dual attention mechanisms, the model effectively captures complex time-series dynamics. Performance evaluation was conducted across two distinct applications. First, on multivariate datasets collected from 70 real-world electric vehicle (EV) trips, TS-MSDA U-Net achieved mean absolute errors within ±1% for key vehicle parameters, including battery state of charge, voltage, mechanical acceleration, and torque. This represents a substantial two-fold improvement over the baseline TS-p2pGAN model, although the dual attention mechanisms contributed only marginal gains over the basic U-Net. Second, the model was applied to high-resolution signal reconstruction using data sampled from low-speed analog-to-digital converters in a protype resonant CLLC half-bridge converter. TS-MSDA U-Net successfully captured non-linear synthetic mappings and enhanced the signal resolution by 36 times, while the basic U-Net failed to reconstruct the signals. These findings collectively highlight the potential of U-Net-inspired transformer architectures for high-fidelity multivariate time-series modeling in both real-world EV scenarios and advanced power electronic systems.