A time-delayed neural network approach to the prediction of the hot metal temperature in a blast furnace

Abstract

The research in this document is motivated by a problem which arises in the steel industry. The problem consists of predicting the temperature of a steel furnace based on the values of several inputs taken one through seven hours in advance (seven different sets of data). Two different time-delayed neural network (TDNN) implementations were used. The data was provided by a large steel plant located outside the United States. This work extends analysis already done by the group on this data using a multi-layer perceptron (MLP). This paper examines the architectures used in detail and then presents the results obtained. A survey of the data mining field related to TDNNs is also included. This survey consists of the theoretical background necessary to understand this kind of neural network, as well as recent progress and innovations involving TDNNs. Issues involved with running computationally intensive neural networks and the optimizations that have led to progress in this domain are also discussed.