With the large-scale deployment of electric vehicle charging piles, their fault prediction has become crucial for grid stability and operational efficiency. This paper proposes a charging pile fault prediction method based on a GBDT-LR (Gradient Boosting Decision Tree-Logistic Regression) hybrid model. The method integrates the nonlinear feature extraction capability of gradient boosting trees with the classification efficiency of logistic regression. This significantly improves prediction accuracy and real-time performance. Firstly, the charging pile fault dataset is preprocessed, including median imputation for missing values and density-based outlier detection, to optimize the quality of input features. Secondly, GBDT is utilized to automatically extract nonlinear features such as voltage harmonic distortion and emergency stop signals, generating high-order feature vectors that are then input into the LR model for classification. Experimental results on a dataset of 85,500 charging piles show that the GBDT-LR model achieves a fault prediction accuracy of 98% and an AUC of 0.97, outperforming traditional models such as SVM (88%) and XGBoost (92%). Additionally, the training time (21.9 s) is reduced by 56% compared to XGBoost (49.8 s). Through the synergistic effect of dynamic feature learning and efficient classification, this model provides a high-precision, low-latency fault early warning solution for the intelligent management of charging piles.