Converter transformer vibration failures

primarily because of winding and core vibrations

severely impact the safe operation of high voltage direct current (HVDC) systems. These vibrations are influenced by voltage and current

and are transmitted through transformer oil and structural components to the tank. Current numerical models fail to accurately capture the vibration transmission process

obscuring the relationship between vibration sources and tank responses. This paper treats the vibration transmission path as a ‘black box’

establishing a database under varying operating conditions based on the correlations among voltage

current and tank vibration. A long-and-short-term time series network (LSTNet) is used to extract local correlations and long-term dependencies

and a dimensional attention mechanism (DAM) optimises input variable weights

whereas the winding and core vibration principles are also embedded in the loss function as boundary constraints. The proposed DAM–LSTNet model is validated under different conditions

demonstrating reliability and portability for complex tank structures. Compared to traditional prediction methods such as long short-term memory (LSTM)

gate recurrent unit (GRU)

convolutional neural network (CNN) and backpropagation neural network (BP)

it shows superior generation performance. The results can expand the dataset of similar vibration characteristics across multiple operating conditions and provide richer sample data for the artificial intelligence algorithm for converter transformer state identification and fault diagnosis.