To predict the Si content of hot metal at No. 2 blast furnace, SSAB, Lulea Works, a three-layer back-propagation network model has been established. The network consists of twenty-eight inputs, six middle nodes and one output and uses a generalised delta rule for training. Different network structures and different training strategies have been tested. A well-functioning network with dynamic updating has been designed. The off-line test and the on-line application results showed that more than 80% of the predictions can match the actual silicon content in hot metal in a normal operation, if the allowable prediction error was set to plus/minus0.05% Si, while the actual fluctuation of the Si content was larger than plus/minus0.10% Si.