Logging data constitutes the basis for oil and gas field development and evaluation. However, in actual mining, factors like poor wellbore stability and equipment failure give rise to the distortion or loss of logging data. A prediction model based on variational mode decomposition (VMD) was proposed to address the issues of unstable and inaccurate results in existing prediction models. The model combines convolutional neural networks (CNN), bidirectional long short term memory (Bi-LSTM), and attention mechanism to predict missing sections in well logging curves. With logging sequence data as input, the VMD algorithm was employed to decompose the sequence into a series of amplitude-modulated and frequency-modulated signal subsequences. The features were extracted by the CNN network and trained by the Bi-LSTM network. During training, the Attention mechanism was utilized to learn the importance weight of each time step dynamically. Finally, the predicted value of the logging curve was outputted. The method was applied to predict logging curves in the Biyang Block of Henan Province and compared with other common machine learning prediction models. The results show that the application effect of the CNN-BiLSTM-Att model improved based on VMD is remarkable, with an error of only the order of 10^-3 and a prediction accuracy of 92.02%. The research results provide new ideas for accurate prediction of logging curves.