Blast furnace is the main equipment in modern ironmaking industry. To achieve hot metal of high quality, the temperature in the blast furnace rector, often indicated by the silicon content in hot metal, should be controlled within a proper range. This brief captures the multiscale features of the blast furnace system and proposes a multiscale predictor based on nonuniform delay-coordinate embedding for silicon prediction. We use the discrete binary particle swarm optimization method to optimize nonuniform embedding parameters, including embedding dimension and time lags. Then, the predictor is trained with the nonuniform embedding features under the framework of the Taylor expansion model. Some experiments on a real blast furnace are conducted to verify the superiority of the proposed method, and the results show that the nonuniform embedding model obtains better performance with lower embedding dimension and higher accuracy than the uniform embedding model, which can be seen as a great improvement of our previous work.