Predictive analysis of the dissolved gas concentration in transformer oil can provide an important basis for its operating status assessment so as to effectively grasp the development trend of the equipment status and provide reference for early warning and maintenance. Given that the existing algorithms are difficult to achieve long-term accurate predictions

a kind of prediction model of dissolved gas concentration in transformer oil based on policy gradient algorithm optimization is proposed firstly in this paper. Firstly

the gate recurrent unit (GRU) is taken as the base

the encoder-decoder structure is introduced to construct the Sequence to Sequence (Seq2Seq) network model

and the Attention mechanism and the Scheduled Sampling algorithm are combined to improve both accuracy and stability of long-term multi-step predictions. Then

the Policy Gradient algorithm in reinforcement learning is used to optimize the hyper-parameters of the network model

and the baseline function and experience pool structure improvement algorithm are added to reduce the strategy network decision variance. The case analysis shows that the method in this paper can deeply extract the relationship between the characteristic parameters time sequence and accurately predict the development trend of the characteristic gas concentration. Compared with the Seq2Seq model

the average relative error and maximum relative error are reduced by 23.91% and 10.22%

respectively. Compared with the LSTM model

the error indicators are reduced by 61.54% and 59.02%

respectively.