Large, interconnected power networks with a high percentage of renewable integration demand a high level of computational power and time for stability analysis. As the analysis needs to be done frequently, development of reduced size equivalent models is required to reduce the computation time. Frequent changes in the operating point can cause significant errors in the reduced dynamic model of the system. Numerous techniques have been proposed to create robust equivalent models that can adapt to such varying operating point, but most of them become unsuccessful in large scale power system applications as the number of possible operating points can be immense. In this paper, a novel long short-term memory recurrent neural network (LSTM-RNN) is proposed to predict the dynamics of a power system while also tracking changes in the operating points. The LSTM-RNN model used in this application is a many to one architecture where boundary measurements are taken as inputs to predict the tie line power flows during dynamic situations. To tackle the model degradation for large variations in the operating point, the Bayesian hyperparameter optimization is used. The proposed model is evaluated using the simplified Australian 14 generator model.