This paper reviews the development history, key technologies, and application value of ensemble prediction, which originated from the understanding of atmospheric nonlinearity and chaotic characteristics. Since Lorenz proposed the “butterfly effect” in the 1960s, it has quantified forecast uncertainty by introducing multiple perturbation experiments into numerical weather prediction. Initial perturbation techniques have evolved from breeding vectors and singular vectors to ensemble Kalman filtering, while model perturbation techniques include stochastic kinetic energy backscatter, stochastic physics parameterization perturbations, and multi-physics ensembles. In the early 1990s, major international meteorological centers successively established global and regional ensemble prediction systems. Through statistical post-processing techniques, ensemble prediction systems have generated various probabilistic forecast products, significantly improving the accuracy and timeliness of extreme weather warnings. In recent years, artificial intelligence (AI)-based ensemble models represented by Google SEEDS etc. have achieved breakthroughs, delivering superior forecast performance at lower computational costs. Future ensemble prediction will develop toward a new paradigm combining physical models with AI to further enhance the forecasting capabilities.