Abstract: The large-scale grid connection of new energy makes the transient frequency response characteristics of new power systems more complex, and the existing online frequency prediction methods make it challenging to balance accuracy and timeliness. Based on this, a transient frequency prediction method based on an adaptive time window driven by data-model fusion is proposed in this paper. Firstly, several frequency curve cyclic prediction models with different length time series data input are trained offline based on a long short-term memory network. Secondly, each power generation cluster's general equivalent frequency response model is established offline using the parameter identification method. Then, a fast analysis model of the system's active power-frequency physical mechanism is constructed. Finally, the frequency curve cyclic prediction model and the active power-frequency physical mechanism rapid analysis model are serial integrated, and a "reliability quantitative evaluation index" is proposed to analyze the accuracy of prediction results at different evaluation moments in the online prediction process in real-time, and adaptively adjust the length of input time series data until the prediction results meet the requirements and output. The simulation results of the IEEE39-node system with wind power show that the proposed method can predict transient frequency response curves quickly and accurately under different wind power permeability or different disturbance and has better evaluation performance than other online prediction methods.