Taking the Weihe River Basin as the research background, the daily runoff data of Xianyang station from 1961 to 2015 are selected as the data input for the Crossformer model. Focusing on the multi-dimensional time series data, this paper adopts the two-stage attention (TSA) mechanism to better capture the correlation between different dimensions. In addition, the Reg-Crossformer model incorporating multi-source covariates is proposed to further enhance the adaptability of the model to complex hydrological conditions. The results of daily runoff prediction in the Weihe River Basin show that compared with the original Crossformer model, the proposed model improves the correlation coefficient (R) and Nash efficiency coefficient (NSE) by 7.46% and 21.63% respectively; reduces the root mean square error (RMSE) by 15.25%. In the comparative experiments of different models, Reg-Crossformer outperforms the conventional machine learning model (SVM) and deep learning models (LSTM and Informer) across various evaluation indicators, demonstrating superior simulation performance and stability. Reg-Crossformer model offers a new approach for the accurate prediction of runoff in the Weihe River Basin, and provides valuable insights into future application of water resources management and deep learning models in hydrological simulation.