To address the issues that the temperature of hot spots inside transformers is difficult to measure directly and the timeliness of traditional simulation calculations is insufficient

and to achieve real-time and accurate perception of the insulation status of equipment

a method for tracing abnormal hot spots in oil-immersed transformers based on sparse neural networks combined with Lévy flight strategy is proposed. Firstly

a simulation model is utilized to calculate and obtain the transformer casing temperature data set under different internal abnormal hot spot parameters. Secondly

based on the simulation dataset

a sparse neural network is adopted to construct a forward proxy model for the transformer temperature field

thereby achieving rapid forward calculation of the temperature field. Finally

this paper proposes an improved crow search algorithm combined with the Lévy flight strategy

which not only enhances the optimization ability of the inversion algorithm but also enables the derivation of internal abnormal temperature point parameters using only a small number of transformer shell temperature measurement points. The experimental results show that

in the transformer abnormal hot spot traceability method

the root mean square error of the temperature calculation of the forward surrogate model is only 0.82. Compared with other traditional network models

the method has a stronger computational fitting ability

which is less than BPNN(1.19)

self attention-BPNN(1.81) and CNN-BPNN(1.84)

respectively. The average location error of abnormal hotspots in the inversion model is 10.8%

and the inversion error of the temperature value of abnormal hotspots is all within 1.8%. Compared with other inversion optimization algorithms

the method is more accurate for tracing the source of abnormal hotspots.