Multi-step prediction of the acetylene dissolved in oil

one of the important discharge degree characterization parameters in power transformer

can provide an important basis for fault diagnosis and early warning of transformer. The existing state prediction models mainly focus on single-step prediction and are insufficient for the prediction means with long time variation trend in the future. In addition

the hyper parameter selection of multi-step predictive model based on deep learning is mostly based on experience and the naive single control variable method

and the coupling relationship between hyper-parameters has not been fully studied. In this paper

a multi-step prediction model of gated recurrent unit (GRU) neural network based on a multi-output strategy is proposed. The coupling relationship of the hyper parameters is studied by changing the structure hyper parameters and training hyper parameters of the model

and the multi-objective gray wolf optimization algorithm is used for hyper parameter optimization of GRU models with different prediction results tendency. The results show that the GRU model can accurately predict the acetylene content in transformer oil for 30 days. The influence of the hyper parameters of the GRU model on the output prediction results is not uniform and affects each other. The same set of hyper parameters cannot achieve multiple goals optimization simultaneously. The multi-objective gray wolf optimization algorithm can optimize the selection of suitable hyper parameters in accordance with the different prediction targets

which provides a reference for the selection of hyper parameters of artificial intelligence algorithms.