Modeling and identification of dynamical systems play a crucial role in simulation and control engineering, the goal is to get an accurate representation of the unknown system. Compared to classical modeling techniques, which require expert knowledge, deep learning approaches are able to learn the dynamics purely from measured input-output data. However, deep learning of sequence-to-sequence models (e.g. recurrent neural networks) lacks proof of stability and robustness often needed in classical engineering. In this study, we propose a hybrid modeling approach that combines a priori knowledge in the form of a linear approximation of the nonlinear ship dynamics with the learning capabilities of a recurrent neural network. The recurrent neural network has constraints on the parameters to get a guaranteed upper bound on the stability gain. By interconnecting the two models, where the constrained recurrent neural network learns the residual of the linear model we achieve stability of the hybrid model, proven using the Small-Gain Theorem. Experimental results demonstrate that the hybrid model outperforms an LSTM model for short-term prediction horizons, while maintaining competitive performance for longer horizons. Keywords: Artificial Intelligence, Recurrent Neural Network, Stability, Hybrid Modelling