The proliferation of distributed generators in the power grid leads to a large number of new types of transient perturbation events that may occur in the system, and the response of the traditional single-equivalent model to the new types of events cannot meet the requirements for safe operation of the power grid. To solve this problem, a set of equivalent models is considered to represent the power grid, each corresponding to a different operating state of the power grid. An adaptive method is proposed to automatically select the appropriate equivalent model under a specific time-level event. System dynamic events are classified as informative or non-informative based on prediction confidence intervals; informative events are used to update the candidate model structure, while non-informative events are used for gray-box model parameter identification. The events are used to drive the model structure selection and updating, and the model correctness is verified by the fit between the predicted and actual dynamics of the model output. The selected and validated models can be used for system dynamic safety assessment. The proposed method is applied to a test system equipped with photovoltaic and wind power and the selection process of the optimal model is given, and the test results verify the effectiveness of the proposed method.