Considering the spatiotemporal dynamics and nonlinear relationships between spatial points, a spatiotemporal extreme learning machine (ELM) is proposed to accurately model time-varying and nonlinear dynamics. First, the nonlinear spatial kernel function (SKF) is developed to describe the nonlinear relationships between spatial points. Then, an online time coefficient model is developed, accounting for the time-varying temporal dynamics. Integrating the SKF with the time coefficient model, the model can adapt to real-time spatiotemporal variation.