Runoff prediction can provide a scientific basis for the rational allocation of water resources. By employing historical hydrological big data as the basis, runoff can be predicted. To effectively utilize time series data such as rainfall and temperature (exogenous variables) to further improve the runoff prediction accuracy, this study proposes a runoff prediction Transformer model based on exogenous variables and sequential embedding. This model introduces endogenous variable sequential embedding in the time attention stage and utilizes the self-attention mechanism to capture the temporal features of sequences and patches, which enhances the model’s ability to capture the temporal features of endogenous variables. Additionally, it distinguishes between endogenous and exogenous variables at the dimension attention stage and employs a cross attention mechanism to capture the influence patterns of exogenous variables on endogenous variables. The experimental results based on the runoff data of the Weihe River Basin show that the proposed approach outperforms the benchmark method in both RMSE and NSE, verifying its effectiveness and practicality in runoff prediction tasks.