Forecasting of blast furnace gas utilisation rate via convolutional autoencoders and K-nearest neighbour methods

Abstract

Gas utilisation rate (GUR) is an essential parameter to characterise the level of energy utilisation efficiency and operation status of blast furnaces (BFs). In the present study, data collected from a BF plant with vanadium–titanium magnetite were subjected to correlation and distribution analyses. Based on the operation cycle of the BF, production parameters from the preceding eight consecutive hours were selected as input parameters for the model. Convolutional autoencoders (CAEs) were employed to extract coded features from the selected data, serving as indirect inputs to the prediction model. Subsequently, K-nearest neighbour (KNN) was utilised to establish a mapping relationship between the coded features and GUR, facilitating forecasting for the subsequent 3 hours. Additionally, a hybrid approach incorporating both supervised and unsupervised training methods was utilised to facilitate GUR prediction, and the performance of the CAE–KNN model was re-validated using data from the remaining sections of the plant. The hit rate of predictions is no less than 96% as the permissible error was within 1.5%.