Abstract: Data-driven machine learning methods have gradually been applied to predict the excavation-induced ground surface settlements.However,the existing neural network models only consider the data at a single monitoring point and ignore the spatial correlation between multiple monitoring points.This study proposes a seq2seq-ConvLSTM2D neural network model to consider the multi-dimensional spatiotemporal features.The model transforms the excavation-induced ground surface settlements into digital images and achieves a long-term settlement prediction at multiple points.The model combines 2D convolutional operation and long short-term memory (LSTM) neural networks to use all the monitoring data around the excavation and to fully exploit the spatiotemporal features of the data.The settlement prediction for multiple days is provided based on a sequence-to-sequence (seq2seq) structure.The effects of input and output sequence lengths are then studied.As the actual monitoring data usually have missing values,a distance-weighted K-nearest neighbors (KNN) model is used to fill in the missing data.The proposed method is illustrated using the excavation project of Zhejiang Conservatory of Music Station.The results show that the determination coefficient (R²),mean absolute error (MAE),and mean squared error (MSE) for settlement predictions are 0.96,0.71mm,and 0.90mm²,respectively.The proposed model outperforms the LSTM,CNN-LSTM,and ConvLSTM1D models in terms of prediction accuracy.The proposed model can provide long-term predictions of excavation-induced settlement,which helps decision-makers better plan and allocate resources to deal with potential engineering problems.