Abstract: To address the problems of insufficient prediction accuracy caused by the complex nonlinearity of hydropower unit vibration signals, as well as the imbalanced capture of temporal and channel features by conventional single prediction models, a vibration trend prediction method based on augmented successive variational mode decomposition (ASVMD) and parallel model is proposed for hydropower units.? First?, population initialization and predation position update strategies are designed to construct an enhanced black?winged kite algorithm integrated with a novel fitness function, which realizes adaptive optimization of key decomposition parameters and obtains multi?dimensional low?complexity decomposed sequences from the raw vibration signals. Second?, considering the high computational cost of the original Informer and the weak feature?modeling performance of the extended long short?term memory network, an adaptive dynamic balance self?attention module and a noise reduction feature learning module are embedded to optimize the model architecture. Base on this, a parallel prediction network is constructed to strengthen the joint capture of temporal and channel features. Experimental results demonstrate that the proposed method achieves the lowest values across four error metrics, the highest fitting degree between predicted and actual values and the narrowest and most stable prediction error fluctuations. These findings fully validate the practicability and high-precision prediction performance of the proposed method, suggesting its broad prospects for engineering applications.