Inversion of the temperature field in oil-immersed reactors using optimal measurement points selected by random forest

To address the subjective issue of selecting measurement points based on mainstream line methods for hotspot temperature inversion in oil-immersed power equipment, this paper demonstrates an oil-immersed reactor temperature field inversion method based on random forest (RF) measurement point optimisation. Firstly, a temperature field calculation method for a 22-kV oil-immersed reactor is proposed. In combination with Latin hypercube sampling, 50 sets of temperature field data are calculated. Based on these samples, the selection of measurement points based on RF feature importance and the training of the genetic algorithm-optimised back propagation (GA-BP) inversion model are undertaken. Finally, the optimal combination of external tank wall measurement points is determined based on comprehensive error indicators, achieving accurate inversion of internal hotspot temperatures in the reactor (with an error of 0.243 °C). The inversion errors are reduced by 2.91 °C and 1.47 °C on average per group compared to existing methods, evincing the superiority of the proposed model.