Accurately predicting structural response under seismic loads is critical for assessing structural damage and performance. Addressing limitations in the predictive capability and efficiency of current seismic response prediction methods for long time series data, as well as considering the scarcity of seismic data, this study adopts a novel neural network model that combines the Kolmogorov-Arnold neural network and gated recurrent unit, along with a multiscale hybrid method. This approach effectively enables real-time prediction of multi-degree-of-freedom response time histories in building structures under conditions of limited seismic data. To verify the accuracy and efficiency of the proposed method, four case studies were conducted on publicly available datasets. Additionally, ablation and comparative experiments were performed to further evaluate the feasibility of the model. Results demonstrate that the proposed method accurately predicts acceleration, velocity, and displacement time histories across multiple degrees of freedom within the structure, surpassing the prediction accuracy of existing models, such as the Long Short-Term Memory (LSTM-f) network and Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model. This model achieves efficient, accurate predictions even with highly limited training data, fulfilling the requirements for practical engineering applications.