<p>Improving accuracy while maintaining low complexity is a crucial challenge in daily streamflow forecasting. To reduce model complexity and computational cost, this paper proposes a high-order lightweight dynamic framework (HoLDF), which can be divided into two parts. First, the identification of high-order structural information of the streamflow system is achieved by improved Granger causality inference approach. Second, the inferred structural information is fed to the dynamic framework for effective training with the original sequence to achieve more accurate predictions. The results show that HoLDF outperforms baselines, reducing RMSE values by 45.33%~86.33% and improving NSE values by 0.233 ~ 0.575, with smaller errors and more timely warnings in peak prediction. Meanwhile, it also possesses higher computational efficiency, with other models requiring more than twice the training parameters and runtime of HoLDF. Therefore, HoLDF exhibits high precision and robustness in daily streamflow prediction while has lower training costs and computational complexity, making it possible for operational deployment.</p>