基于自监督预训练与时序注意力机制的变压器顶层油温预测OA北大核心CSTPCD

Transformer fault early warning relies on more accurate and reliable prediction of transformer top oil temperature.The self-supervised pre-training method is used to transform the oil temperature prediction task in the forecasting model training process into the oil temperature reconstruction task,which generalizes the model training method and improves the model's ability to extract historical oil temperature situation information.A dual-channel pre-trained time-series attention network(DPAnet)model is proposed.The model uses a time-series attention mechanism and a deep neural network with trend cycle branches,respectively targeting the trend laws of oil temperature data.The periodic law realizes modeling at the single time step level,thus enhancing the predictive ability on multiple time steps.The example analysis shows that in the short-term prediction scenario of 1～72 h with hourly granularity,the average prediction loss of the top oil temperature prediction model proposed in this paper is 1.847,and the average certainty coefficient is 0.862,which improves the prediction compared with other models.The model has strong generalization ability and robustness,and effectively supports the analysis of the change trend of oil temperature on the top layer of the transformer.