Abstract: To address the issues of noise interference in actual monitoring data of dissolved gases in transformer oil and the ineffective extraction of deep features between data by a single long short-term memory (LSTM) neural network, a combined prediction model combining correlation variational mode decomposition (CVMD), 1D convolutional neural network (1D-CNN) and LSTM is proposed. Firstly, CVMD is used to remove the noise signal from the original gas sequence, and the denoising sequence is decomposed into a group of relatively stable subsequence components. Then, CNN-LSTM prediction models are constructed for each subsequence component, and 1D-CNN is used to mine deep features between data to form feature vectors, which is input into LSTM for prediction. Finally, the prediction results of each sub-sequence is superimposed and reconstructed to obtain the final gas predicted value. Four groups of comparison experiments were carried out to verify the proposed model in a comprehensive and multi-angle way. The results of the example study show that the average absolute percentage errors of single-step and multi-step prediction of the proposed model are 1.53% and 2.09%, respectively. Compared with the existing model, the proposed model in this paper has significantly improved the performance of the single-step and multi-step prediction. The study provides important technical support for transformer on-line monitoring and fault warning.