To more accurately predict the surrounding ground surface settlement resulting from unloading during foundation pit excavation, a neural network algorithm with feedback mechanism and reverse learning (FBRLNNA) was proposed in this study, which was based on the neural network algorithm (NNA). By integrating this with back propagation (BP) neural networks, a surface settlement prediction model was constructed and subsequently applied and validated in the foundation pit of Qingdao Metro Line 15. In comparative tests with nine optimization algorithms across 18 benchmark functions, superior optimization results were achieved by FBRLNNA in 80% of the cases. Additionally, an analysis of the FBRLNNA-BP model against four other settlement prediction models shows that it performs best in terms of mean squared error (MSE), mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²), with prediction errors maintained within 5%, demonstrating the model's enhanced accuracy for settlement prediction. The research results provide a novel approach and valuable reference for predicting surface settlement induced by foundation pit excavation.