Transformer is the core equipment in power system and its safe operation is the foundation of the stability and reliability of power system.Prediction of gas contents and the development tendency based on on-line monitoring historical data is the key to DGA failure prediction the necessary supplement to on-line monitoring and an important means of failure prediction of transformers.The existing algorithms have two major drawbacks:①the datasets are too small to train and validate the proposed models;②most previous methods make an assumption of that samples from one transformer are independently and identically distributed while paying less attention to the time correlation between samples.To increase predictive ability this paper studied the supervised learning and sequential learning methods and their applications in data prediction of gas dissolved in transformer oil.Then a large-scale dataset was built and released to provide a comprehensive validation set and increase the generalization.Lastly a walk-forward split method was adopted to augment the data.Extensive experiments conducted on a large-scale dataset have well validated the competitiveness of sequential learning on data prediction of gas dissolved in transformer oil.