Based on multi-feature fusion (MFF) and Newton-Raphson-based optimizer (NRBO), a long short-term memory (LSTM) model was proposed to improve the accuracy of daily runoff prediction and address the limitations of existing models in extracting multi-feature data. Variational mode decomposition (VMD) was applied to decompose the original runoff series into several intrinsic mode functions (IMFs) in the spectral domain, reducing the non-stationarity and complexity of the runoff series. Hydro-meteological features highly correlated with runoff were selected by correlation analysis. These selected features were integrated with IMF components, which could be used to construct the feature matrix serving as the model input. The NRBO algorithm was utilized to optimize the key hyperparameters of the LSTM model to achieve the optimal configuration. Datasets from Huayuankou and Lijin stations in the Yellow River Basin were selected for experiments to validate the model performance. Results showed that for the novel model, Nash-Sutcliffe efficiency (NSE) values at two stations were all higher than 0.97 under forecast horizons of 1, 2, and 3 days. Compared with Random Forest, CNN, TCN and other models, the proposed model —— by fully fusing multi-feature data —— achieved an average 61.42% reduction in RMSE and 4.20% increase in NSE over a 1-day forecast horizon at two stations, effectively improving the daily runoff prediction accuracy.