ASFC-based DNN Modeling for Prediction of Silicon Content in Blast Furnace Ironmaking

Abstract

The real-time and accurate prediction of the silicon content in the blast furnace (BF) plays an important role in controlling the temperature of the BF and stabilizing the BF condition. As the BF system model continuously changes during the smelting process, a single model may not be able to fully represent the complex operating conditions of the BF. For this reason, this paper establishes a distributed neural network model based on online adaptive semi-fuzzy clustering algorithm (ASFC-DNN) to predict the silicon content in molten iron in hope of reasonable BF control guidance. The model adopts an online self-adaptive semi-fuzzy clustering learning algorithm (ASFC) which can reflect the fuzzy clustering characteristics of BF production data. Through the hybrid clustering of production data, a distributed neural network sub-model is established for the decomposition of complex conditions. ASFC also updates the production data category to update the network parameters of the distributed neural network, and the reliable network structure for the current input production data is obtained to predict the silicon content. The performance of the ASFC-DNN is compared with the BPNN and ordinary elmanNN, and the ASF C-DNN’s hit rate and RMSE are better than that of other models. The study found that the ASFC-DNN model has more stable prediction errors and higher prediction accuracy than the above models. On-site experiments demonstrate that the ASFC-DNN model has higher accuracy and better tracking performance and it provides reliable silicon content for operators.