A Dynamic-Response-Enhanced Active Current Prediction Method for Synchronous Reluctance Motors Under Multi-Operating-Condition Switching

Synchronous reluctance motors in quadruped robot joint drives are prone to active-current peaks, abrupt rate variations, and switching-neighborhood error concentration under foot–ground impacts and obstacle-induced load steps, leading to prediction lag and peak underestimation. To address these issues, this paper proposes a dynamic-response-enhanced multi-condition active-current prediction method based on TPE-VMD-BiLSTM-TRC. First, the original sequence is segmented according to operating-condition boundaries, and prediction samples are constructed within each segment to reduce cross-condition information leakage and distribution inconsistency. Second, variational mode decomposition is performed on each segmented sequence to separate multi-scale fluctuation components, and BiLSTM is used for mode-wise one-step prediction. Third, TPE jointly optimizes the VMD decomposition parameters and BiLSTM hyperparameters to improve parameter matching under different operating conditions. Furthermore, a switching-aware composite loss and a switch-gated residual correction branch are introduced to enhance dynamic tracking and compensate for structural bias in switching neighborhoods. Experiments on variable-frequency drive data show that, compared with TPE-VMD-BiLSTM, the proposed method reduces the overall RMSE, switching-neighborhood RMSE, and MAPE by approximately 29.8%, 38.8%, and 7.4%, respectively. On the additional held-out operating-condition segment, the overall RMSE is reduced by 28.8%, indicating stable prediction performance across different segments.