To capture the dynamic variation laws of hydrological processes and improve runoff prediction accuracy, by combining three kinds of correlation analysis with covariance analysis, we systematically analyzed the correlation characteristics between different meteorological variables and runoff, dynamically extracted high-contribution features for runoff prediction and optimized their weights. Then, based on the deep temporal convolutional network(DeepTCN), a p DeepTCN model for multi feature runoff prediction was constructed to capture both the long term dependencies and short term fluctuations in runoff sequence. The measured data of the Atsuma River Basin in Hokkaido from 2015 to 2020 were used to verify the pDeepTCN model, and the results show that the p DeepTCN model achieves the optimal prediction performance compared with other deep learning models. Specifically, the Nash Sutcliffe efficiency coefficient of the pDeepTCN model is increased by 28.48%, and the root mean square error is reduced by 27.63%. Although its percent bias is 23.34% higher than that of the WPMixer model, it is 17.44% lower than that of the other deep learning models. Meanwhile, although its coefficient of determination is 0.011 lower than that of the original DeepTCN model, it is 10.17% higher than that of other deep learning models. In summary, the pDeepTCN model significantly improves the accuracy of runoff prediction, and exhibits superior adaptability and robustness.