Data-driven neural network methods have been widely applied for the prediction of stress–strain behavior, but have proven ill-suited for the extrapolation of time-dependent creep behavior. To overcome this problem, we embedded a physics-based model into feedforward neural networks (FFNNs) to construct a model-guided neural network (MGNN). We proposed a new initialization method for the weights in the model, based on selecting the appropriate physics-based model and activation function, and the resulting MGNN was used for predicting the creep behavior of blade-grooves in a steam turbine rotor under in-service conditions. We compared the performance of the MGNN with baseline methods, namely MGNN0, a FFNN, and a nonlinear autoregressive network with exogenous inputs (network). The results showed that the physics-based model and the neural network in the MGNN complemented each other: the model provided physical relationships to guide the neural network, and the neural network provided stress-fluctuation-tracking for the model. This functionality enabled primary creep behavior to be used as training data for the MGNN, enabling it to predict secondary creep behavior.