STEEL PLANT MODELING AND ANALYSIS USING A NEURAL NETWORK
Jürgen Van Gorp · 1999
This paper is structured as follows. The steel plant is described and the measurement data is analysed in section II. The next section of the paper describes how a Neural Network is chosen and trained as a black box model for the steel plant. It is shown that the NN has a better prediction performance than the currently used manual predictions, and that an upper limit exists for the performance. Section IV gives an analysis of the measurement data, based on the obtained NN. Suggestions are given where to improve the measurements in order to improve the NN performance. II. PLANT DESCRIPTION The steel is processed in two phases. In a first stage raw iron and waste steel are melted with added chemicals. The melted iron is then poured in a convertor and during the second stage pure oxygen is blown into the melt to obtain a specified Carbon content and specified goal temperature. This two stage process will further be called a batch. The total number of input and output parameters for the system equals 59. Not all of these parameters are useful for the NN modeling and a selection was made with a maximum of 25 input parameters, based on expert knowledge. Table 1 gives an overview of these parameters, with their tolerances and minimum/maximum boundaries. An (I) indicates that the value is known prior to the batch and can be used as an input to the NN. An (O) indicates that the value is measured at the end of the batch and can only be used for validation. The sensitivity values and are discussed in detail in section IV. In the sequel references to the parameters in table 1 will be written in italic.