To address the problem of uncertainty of prediction factors and model complexity of traditional runoff prediction methods, prediction factors were selected based on feature importance analysis of monthly runoff time series, and the nonlinear relationship between runoff time series was captured by the mixed kernel function-support vector machine (MKSVM) model. An improved grey wolf optimizer (IGWO) that integrated multiple strategies, such as dynamic lens imaging reverse learning and Lévy flying strategies, was proposed to enhance the stability of the global parameter optimization of the MKSVM model, and an IGWOMKSVM model for runoff prediction was constructed. The results of monthly runoff prediction at Yingluoxia Hydrological Station in the Heihe River Basin show that the NashSutcliffe efficiency coefficient, root mean squared error, and KlingGupta efficiency coefficient of prediction results of the IGWOMKSVM model were 0.8942, 16.9099 m3/s, and 0.8639, respectively. Compared with the traditional SVM model, the IGWOMKSVM model has high adaptability in runoff prediction, and compared with the long shortterm memory network model and the seasonal autoregressive integrated moving average model, the IGWOMKSVM model can better predict the real change process of monthly runoff.