Abstract: Case-based reasoning was used to predict the starting temperature of molten steel in second refining so as to avoid the long training time of a BP (back propagation) neural network. Analytic hierarchy process (AHP) was applied to determine the weights of factors influencing the starting temperature. Grey relational degree was adopted to compute the similarity between cases. Thus the shortcoming of difficulty in obtaining accurate cases with incomplete information is conquered. A four-step search method, including class search, rough search, delicate search, and optimized search, was provided, by which the search time decreases greatly. Experimental results using both artificial neural networks and case-based reasoning were compared. It is shown that case-based reasoning has got a higher hit rate and a shorter response time than artificial neural networks.