Power Transformer Health State Prediction Based on Multi-attention Coupling Learning

Transformer is the key apparatus of the power system, and its operational reliability is directly related to the stability of the power grid. Accurately predicting the future health state of transformers can reserve enough time for formulating outage maintenance and power dispatching plans. The existing transformer health state prediction methods are based on the structure of “predict-diagnose”, which is the superposition of two local optimal methods, with error superposition. To fill this gap, a transformer health state prediction method based on multi-attention coupling learning is proposed in this paper. The “situational quantity” integrating the transformer operation state change relationship, the correlation relationship between state quantities, the mapping relationship between the state quantity and the state, the fuzzy relationship between the evaluation result and the actual state, and the traditional expert experience is proposed. Then, the structural optimization attention (SOA) model applicable to “situational quantity” is proposed, and various relationships are treated as individual attention. All kinds of attention are coupled and constrained in the SOA to realize the unification of the prediction process and diagnosis process, obtain the global optimization of prediction and diagnosis, and the problem of error superposition is solved. The actual fault case verification results show that the accuracy of the proposed SOA prediction model can reach 96.33%, which is significantly better than the traditional methods.