Abstract: 　　In the process of processing data using a fixed-size sliding window, it is difficult to dynamically extract time-frequency features based on differences in ground motion intensity and amplitude variations of structural seismic responses, which can easily lead to the loss of key features and a reduction in prediction accuracy. This paper proposes a structural response prediction method that uses a variable-size sliding window for data processing. Taking a frame structure as the research object, and using 100 ground motion acceleration time-history records and the inter-story drift ratios output by a finite element model as the dataset, the proposed algorithm predicts the inter-story drift ratios. Comparison with prediction results based on a fixed-size sliding window shows that the proposed algorithm can effectively predict the response of the frame structure. Compared with the fixed-size sliding window method, the mean absolute percentage error is reduced by 12.27%, and the peak inter-story drift ratio error is reduced by 7.87%.