<p>During construction, the deformation of underground engineering support structures in large-scale excavations significantly affects overall structural safety. However, traditional deformation prediction methods often have limitations, such as overoptimized assumptions, complex model parameters, and difficulty in reproducing tests. To improve the mining depth and prediction accuracy of deformation data for underground engineering support structures, we first consider the spatio-temporal (ST) correlation of the deformation data. We introduce an expansive convolutional neural network (CNN) to analyze the spatio-temporal matrices of the deformation data. This allows us to perform spatio-temporal feature extraction and analysis by combining multiple convolution branches with a long short-term memory (LSTM) neural network. Next, we apply two different attention mechanism algorithms to determine the feature weights and deep learning of the target prediction features. This leads to the construction of an ST-CNN-LSTM deformation prediction model tailored for large-area excavation underground engineering support structures. Finally, we analyze the subsequent development law of the overall deformation of underground engineering support structures using the constructed ST-CNN-LSTM model. This analysis is based on three large-scale underground excavation projects in Beijing, China. We also perform a comparative analysis between the complete ST-CNN-LSTM model and its variants, including ST-CNN-LSTM(sin), ST-CNN-LSTM(nse), and ST-CNN-LSTM(s–n). The comparison focuses on deformation prediction accuracy and error evaluation indices. The results show that the error of the complete ST-CNN-LSTM model in deformation prediction is within 10% of the measured value. The prediction accuracy is significantly higher than that of the other models, particularly during long construction periods. The ST-CNN-LSTM model can more accurately capture the development trend of overall deformation in support structures.</p>