This study evaluates the effectiveness of the Dynamic Nelson-Siegel (DNS) model in forecasting Chinese government bond yields using daily zero-coupon yields from March 2006 to April 2024. We optimized the DNS model's decay parameter λ to better fit the Chinese market. The DNS model provides a simple yet robust representation, effectively capturing factor evolution with an AR (1) specification. Additionally, we explore machine learning methods to enhance yield curve forecasting. We apply eXtreme Gradient Boosting (XGBoost) to address the complex, non-linear patterns in yield curves and Long Short-Term Memory (LSTM) networks to capture temporal dependencies in the data set. After tuning the network hyper-parameters, our results show that both machine learning models slightly outperform the DNS model. However, the modest improvements highlight the continued value of the DNS model's simplicity and interpretability in yield curve modeling.