Before a transformer fails, the gas content in transformer oil exhibits certain regularity over time. Therefore, predicting the content of fault characteristic gases dissolved in transformer oil can help predict the type of fault. However, monitoring data for dissolved gases in transformer oil often suffer from noise, outliers, and nonlinearity, making traditional statistical analysis methods less effective. This paper proposes a combined prediction model called CEEMDAN-CNN-LSTM, which integrates Complete Ensemble Empirical Mode Decomposition (CEEMDAN), Convolutional Neural Networks (CNN), and Long Short-Term Memory networks (LSTM) to address the prediction issues in time-series data of dissolved gases in oil. First, density clustering algorithms are used to detect and clean anomalies in the dissolved gas data, removing noisy data. Then, the CEEMDAN method is applied to decompose the time-series data of dissolved gases in oil into multiple modal components, thereby enhancing the predictability of the data. Next, CNNs are employed to extract features from each modal component, and the LSTM network is used for prediction and reconstruction to obtain the final results. The paper will detail the theoretical foundation and implementation process of the proposed method, and through comparative analysis of prediction results under different algorithms, it will demonstrate the superiority of this model in predicting dissolved gases in oil.