Abstract: Due to the influence of complex geological conditions and dynamic construction environments on the deformation development trend of deep foundation pits, accurately predicting the deformation of deep foundation pits to ensure their safety is an urgent engineering challenge that needs to be addressed. To achieve high-precision and strong robustness in deformation prediction, this study proposes a hybrid deep learning model (TCN-LSTM) that integrates Temporal Convolutional Networks (TCN) and Long Short-Term Memory Networks (LSTM). This model extracts multi-scale temporal features through the dilated causal convolution operation in TCN, and utilizes the gating mechanism of LSTM to model long-term dependencies and nonlinear dynamic evolution processes. The cross-modal integration of TCN and LSTM effectively enhances the model's feature representation and generalization capabilities. Comparative experiments based on measured data from a deep foundation pit project at a hospital in Guangzhou show that the proposed TCN-LSTM model significantly outperforms traditional RNN, LSTM, and CNN-LSTM models in terms of goodness of fit (R²), mean squared error (MSE), and mean absolute error (MAE). Specifically, R² increases by 97.43%, 80.59%, and 11.38%, respectively; MSE decreases by 33.01%, 23.66%, and 10.13%, respectively; and MAE decreases by 57.81%, 49.00%, and 35.46%, respectively. Additionally, the model exhibits excellent noise robustness. This study provides a reliable solution for deep foundation pit deformation prediction and holds significant theoretical value and engineering application prospects for intelligent perception and proactive prevention and control of engineering risks.