Abstarct: Zinc oxide surge arresters are critical components in power networks, serving to protect equipment from lightning strikes. However, during normal operation, these surge arresters remain inactive and are susceptible to leakage currents, which are influenced by environmental conditions and the structural properties of the surge arrester. As a result, precise and effective monitoring of surge arrester performance through non-destructive methods, such as leakage current measurement, is of significant importance. This study introduces a novel approach for modeling the leakage current of zinc oxide surge arresters under varying environmental conditions. The model utilizes artificial intelligence, specifically the Random Forest method, to estimate the electrical conductivity of the contaminated laxyer. The primary objective of this research is to replace the direct estimation of leakage current with the estimation of the electrical conductivity of the contaminated laxyer, in order to achieve a more comprehensive and scalable model. To analyze the results, three scenarios were defined. In the first scenario, leakage current was derived baxsed on the electrical conductivity obtained from laboratory tests under various conditions, and the results were compared with experimental data. In the second scenario, leakage current was calculated using the Finite Element Method (FEM) model, where the conductivity was estimated by the Random Forest algorithm. In the third scenario, leakage current was extracted from the FEM model using conductivity predicted by the Random Forest method. The mean squared error (MSE) for the Random Forest method was 0.000216, which is significantly lower than other methods. Additionally, the R² value of 0.9999 indicates an excellent fit between the model's predictions and actual data in estimating the electrical conductivity of the contaminated laxyer. The results not only matched the waveform of the leakage current but also showed high accuracy in the amplitude of harmonic components. In the third scenario, despite not utilizing some experimental data for training, the model's results demonstrated that the maximum error in the harmonic components of the leakage current was less than 10%, underscoring the model's strong accuracy even in cases where experimental data was not directly used for training.