Abstract: The existing genetic algorithm (GA)back propagation (BP) neural network prediction model is inefficient in pretreating training samples and the structural design of hidden layers. The bestnormalized interval is determined using correlation coefficient regression analysis, the analytic expression of the number of neurons in the hidden layer is derived using the principle of statistics, and the corresponding range of the best number of neurons in a single hidden layer is proposed. The results show the following: (1) The bestnormalized interval of the training sample pretreatment is ［0.05, 0.95］. (2) The best number of neurons of a single hidden layer structure is in the interval of ［4, 7］, and a double hidden layer structure is more suitable for the case of neurons （> 4）. (3) The metabolism approach of selecting training samples can significantly reduce the timespace effect and human interference of foundation pit deformation. (4) The 42 (7, 8)1 GABP neural network model is constructed to predict the horizontal displacement of foundation pits with different depths, and the accuracy evaluation index shows that the prediction effect of the optimized GABP model is good, which provides references for engineering development. By comparing the optimized model with other models, its superior effect is validated.

Key words: foundation pit engineering, genetic algorithm-back propagation neural network, foundation pit monitoring, horizontal displacement, deformation prediction