Distribution of basic oxygen furnace (BOF) flame is an important basis for determining the content of molten steel temperature and composition. Analyzing 350~1100 nm spectral data from the furnace mouth, furnace flame atomic emission spectra overlap in a continuous or "black body" radiation, which are in a clear visible radiation. Data collected from nanjing iron and steel company′s steel-making furnace as sample data are used to implement the algorithms. The sample contains 400 data pairs. A model is applied based on the theory of wavelet analysis and neural networks to predict the temperature of the furnace flame and the results are analyzed in detail. It is shown that the method of neural networks with compact structure can give better hit rates of prediction; the temperature predicted by the model is inosculated to the temperature obtained by converter sub-lance comparatively.