Adaptive triple-phase-shift control of a dual-active-bridge converter for 800-V AI rack buses

Purpose This paper proposes a load-profile-aware adaptive triple-phase-shift (LA-ATPS) control strategy for dual-active-bridge converters in 800-V AI rack buses under non-unity voltage gain, rapid load-envelope transients and constant-power-load instability. The method aims to simultaneously reduce leakage-current stress, preserve transient current continuity during TPS mode transitions, and enhance bus-side damping without disturbing efficient operating points. By separating steady-state optimization, transient waveform migration and CPL-oriented impedance shaping within a unified TPS control surface, the proposed approach improves both dynamic stability and power-transfer performance in high-density AI data-center applications. Design/methodology/approach The proposed controller decomposes the TPS command into three coordinated layers: steady-state current-stress minimization, one-cycle transient current-state matching and sensitivity-normalized virtual impedance injection. A hybrid event classifier activates transient control only during fast load-envelope variations. Steady-state TPS operating points are generated offline and stored with local phase-sensitivity information for online damping normalization. A one-cycle transition constraint verifies waveform feasibility through endpoint-current, flux-bias, pulse-width and ZVS conditions before PWM updating. The approach is validated using switching-level simulations and a 5 kW experimental DAB prototype operating at an 800V rack bus. Findings Results demonstrate that the proposed LA-ATPS strategy improves both transient compatibility and CPL stability under demanding AI rack-bus operating conditions. Simulations show up to 13% leakage-current RMS reduction at non-unity gain, a decrease in CPL-induced bus-voltage ripple from 7.30–5.07 Vp-p and a 14.4% reduction in line-current fluctuation during periodic load perturbations. Experimental results verify shorter post-step ringing, bounded bus-voltage recovery and balanced leakage-current waveforms after TPS transitions. The controller effectively prevents accumulated flux bias and suppresses prolonged oscillation without introducing excessive damping or degrading steady-state efficiency. Research limitations/implications The proposed method requires detailed offline TPS characterization and exhaustive firmware-level mode-boundary verification to ensure safe interpolation across operating regions. Virtual-impedance performance also depends on accurate plant partitioning and damping-band selection. Although the present study validates transient behavior and waveform compatibility, further work is required to establish dense efficiency mapping and calibrated steady-state measurements over the full operating envelope. The study nevertheless demonstrates that modulation feasibility, transient state migration and CPL stabilization should be jointly considered in future high-voltage AI data-center converter design rather than treated as isolated control objectives. Originality/value This work introduces a unified TPS control framework that explicitly separates steady-state optimization, transient current-state compatibility and CPL-facing impedance shaping within the same modulation surface. Unlike conventional TPS methods focused only on efficiency or transition continuity, the proposed strategy embeds waveform feasibility and sensitivity-normalized damping directly into TPS command generation. The method further incorporates local power-to-phase sensitivity into virtual-impedance control, enabling operating-point-consistent damping across varying gain conditions. The study provides a practical control architecture for 800-V AI rack buses and demonstrates how TPS degrees of freedom can be physically allocated to multiple converter objectives without compromising stability or efficiency.