In the study of deformation characteristics of deep foundation pit excavation in a soft soil environment， the hardening elastic-plastic model is often used for analysis， such as the HSS model and MCC model. In the soft soil area of the Nanjing River floodplain， large local deformation often occurs during deep foundation pit excavation. Some soil deformation states are between small and large strains， so a single model cannot accurately predict the deformation characteristics of soil. At the same time， the BP neural network has been widely used for predicting foundation pit deformation prediction. However， in the training process， the weight threshold easily falls into the local optimal solution， which affects the accuracy of prediction. Based on this， relying on the typical soft soil deep foundation pit project in the Nanjing area， the HSS model and MCC model in Midas are used to analyze the difference in pile deformation between the two models， and the two models are linearly fused based on the least squares criterion. The fusion model can calibrate and supplement the monitoring data of the subsequent section. The BP neural network is optimized by fusing the sparrow search algorithm， and the global optimal weight threshold is obtained by fast convergence in the training process. Based on the monitoring data of the excavated section of the narrow and long foundation pit， the training is learned. The deep deformation characteristics are revealed according to the shallow excavation of the subsequent section. The predicted results are in good agreement with the measured values. The research results have important reference value for predicting the large deformation of deep foundation pits in soft soil areas.