Accurate prediction of wall deflection induced by braced excavations is of great importance to prevent potential damage to retaining structures and the surrounding environment. Although previous studies have utilized spatiotemporal deep learning algorithms for this purpose, they overlooked the underlying deflection mechanism when developing these algorithms. To address this limitation, this paper proposes a novel physics-guided architecture of neural networks (PGA-NN) for predicting wall deflection where the physical mechanism is hardcoded in the neural network architecture. In the PGA-NN, a physical intermediate layer and a monotonicity-preserving long short-term memory (LSTM) cell are integrated through understanding the underlying physical mechanism of wall deflection. The performance of the proposed PGA-NN model was verified using data from an excavation project. Results indicate a significant superiority of the proposed PGA-NN model over the baseline model. The PGA-NN model demonstrates a strong capability for accurately forecasting wall deflections in advance by considering the physical mechanism.