In complex operating environments of power transformers, the dissolved gases in transformers have non-stationary and nonlinear characteristics. The prediction models of the neural network are difficult to meet high accuracy and reliability requirements which only consider the temporal features. During the data collection process, it is inevitable to exist outliers, which leads to a decrease in data quality and subsequently affects the accuracy of the prediction model. Firstly, density-based spatial clustering of applications with noise (DBSCAN) is proposed to clean the time-series data of dissolved gases in oil in this paper. Then, the adaptive nonlinear weight and Levy flight strategy are proposed to improve the whale optimization algorithm, enhancing its local and global optimization capabilities. The improved whale optimization algorithm is used to optimize hyperparameters in DBSCAN which improves the efficiency of data cleaning. Finally, the complex correlation between gases is analyzed, and a spatiotemporal coupled convolutional neural network model is constructed to mine the spatiotemporal characteristics of gases and achieve gas prediction. Verified by the dissolved gases in the oil of the power station, the results show that the R-squared increased by 0.727 after data cleaning. The R-squared is above 0.9 in all six characteristic gas predictions. Compared with other models, this prediction model proposed in this paper has achieved the best prediction results in feature gas prediction, which demonstrates the effectiveness of the prediction models.