Based on the characteristics of shale gas wellbore stability being affected by multiple nonlinear coupling factors, an improved BP neural network prediction model was constructed. On the basis of the traditional BP structure, this model introduces three hidden layers and ReLU activation function, combined with Batch Normalization and Dropout to suppress overfitting, and adopts focus loss and AdamW optimization strategies to improve the recognition ability of imbalanced samples. Using measured downhole data to construct a training set, the results showed that the F1 score of the conventional well section reached 0.86. The experiment shows that the model maintains good robustness under high structural heterogeneity conditions and has engineering application value.