Reliable prediction of the initial tunnel support and surrounding rock stresses is essential for tunnel construction to ensure structural stability, sustainable operation, improve safety and reduce maintenance costs. However, due to big data, the prediction of work is difficult to conduct using traditional means. Therefore, a backpropagation (BP) neural network method adopted in the study to enhance the prediction process for predicting the initial support and surrounding rock stresses caused by Yangjiashan mega-section tunnel construction. The BP neural network is a type of artificial neural network that uses a supervised learning algorithm to train the network. In order to determine the best model, this study adopted the five different BP neural network methods. For each gradient descent method, each network was trained by using 912 sets of measured data (a total of 5472 data) to analyse the differences between the predicted and true values of stresses and use to analyse the distribution of prediction errors. A comprehensive evaluation of the five training methods was carried out based on the analysis of a multi-objective optimisation problem. The results show that, in order from best to worst: trainscg > traingdx > traingrp > traingda > traingdm. Trainscg reported prediction accuracy of 93.8%. Hence, the BP neural network by trainscg is used in the subsequent project to predict the initial support and surrounding rock stresses generated by the subsequent tunnel excavation. To sum up, the BP neural network can be effectively used to predict the initial support and surrounding rock stresses arising from the subsequent tunnel excavation.