The deformation of retaining structures during foundation pit construction is influenced by multiple factors, including excavation sequence, strut installation, soil properties, and construction time effects. Such deformation exhibits obvious dynamic evolution characteristics. When traditional elastic foundation beam methods are used for prediction, soil parameters and effective strut stiffness cannot be updated in real time with construction processes and monitoring feedback. When purely data-driven neural networks are used for prediction, the predictions are limited by insufficient physical constraints and limited training samples. To balance the mechanical interpretability of the elastic foundation beam method and the parameter optimization capability of machine learning, a physics-informed machine learning prediction model that embeds the forward computation of an elastic foundation beam finite element is proposed in this paper. The model took the elastic foundation beam method as its physical basis. The forward mapping relationship among the horizontal proportionality coefficient of subgrade reaction, effective strut stiffness, and the displacement response of the retaining structure was established through the finite element method. This finite element computation process was then embedded into the computational graph of machine learning. During model training, the horizontal proportionality coefficient and effective strut stiffness were treated as trainable parameters. The mean squared error between the field-monitored displacement and the model-calculated displacement was used to construct the loss function. Key parameters were dynamically inverted, and retaining structure deformation was predicted through automatic differentiation and the backpropagation algorithm. The proposed model introduces the mechanical constraints of the elastic foundation beam method into the machine learning computational graph and training process. In this way, the physical consistency and interpretability of the prediction results are improved, and the dependence of the model on large-scale monitoring data is reduced. Results from an engineering case study show that the model can reasonably reflect the dynamic evolution law of retaining structure deformation in foundation pit construction. The inverted parameters and predicted displacements exhibit good rationality. The model can provide technical support for deformation warning of retaining structures, adjustment of support parameters, and safety control of foundation pits during construction.