Prediction for silicon content in molten iron using unsupervised optimal fuzzy clustering

Hard hierarchical clustering methods are well known methods for recursively partitioning sets to subsets, a data pattern that has been classified to one of the clusters cannot be reclassified to other clusters. But the switches from one stationary state to another are usually vague, so such switches are naturally treated by means of fuzzy clustering. A unsupervised optimal fuzzy clustering approach is established in this paper for predicting silicon content in molten iron which collected online from No.7 BF at Handan Iron and Steel Group Co.. This new approach consists of 4 steps: step 1 Establishes temporal patterns of silicon content ([Si]) time-series and Cluster the temporal patterns into an optimal number of fuzzy sets; step 2 groups similar temporal patterns together into clusters, by an unsupervised fuzzy clustering procedure; step 3 Fits a prediction model (AR) to each cluster; step 4 predicts the future samples of [Si] by a fuzzy mixture of the above prediction models. The new algorithm was applied to predict [Si] only using the last [Si] time series, and good performance is shown due to the high percentage of prediction hitting the target.