Prediction of loss on ignition of light-burned magnesia during flash calcination of magnesite in a transport bed based on feature importance ranking and Bayesian optimization

Abstract Flash calcination of magnesite in a transport bed offers notable advantages. However, the backwardness of testing technology seriously affects the stability of the quality of its product, light-burned magnesia (LBM). This paper addresses the problem of predicting the loss on ignition (LOI) of LBM by proposing a prediction method based on feature importance ranking and LSTM networks. First, the XGBoost model is employed to rank the importance of input features, and a split node is set to filter out key features. Subsequently, an LSTM model is constructed based on the selected features to predict the LOI. To further enhance model performance, a Bayesian optimization approach is adopted to optimize the structural parameters of the LSTM model as well as the feature split node. Experimental results demonstrate that the proposed method significantly improves the accuracy of LOI prediction, providing robust support for optimizing the flash calcination process of magnesite in a transport bed.