Long short term comprehensive prediction of sinter FeO components based on EEMD and machine learning

Iron making depended on "70% of raw materials, 30% of operation". Sinter was the main raw material for blast furnace iron making, and FeO composition was an important index affecting the reduction, strength and particle size of sinter ore, as well as an important factor affecting blast furnace iron production and fuel ratio.Therefore, timely and accurate FeO content of sinter ore played a significant role in guiding the smooth production of blast furnace ironmaking. To tackle the problems of delayed detection results and poor accuracy of FeO composition of sinter ore, an comprehensive long and short-term prediction model of FeO composition by EEMD and machine learning was proposed and established. Exploratory analysis was conducted for sintering data, and according to the characteristics of sintering data, box line diagrams and sliding windows were used to process the data to ensure the data value and to consolidate the data foundation for modeling. The comprehensive model consists of two modules. Long-term prediction model. Applying EEMD (ensemble empirical mode decomposition)to decompose fluctuating FeO component data, which could reduced the complexity of input data, and perform advance forecasting of FeO components within 3 h with Bi-LSTM(bi-directional long short-term memory). Short-term prediction module. Integrating EEMD, feature selection and extraction methods to construct derived features and enhance the learning ability of the model for input and target data to ET(extra-trees) the FeO composition prediction for the next hour. Under the validation of the unknown sintered data test set, it was found that EEMD-assisted machine learning modeling could significantly improve the accuracy and stability of FeO composition prediction, and the MAPEand MSE of EEMD-Bi-LSTM and EEMD-ET models were around 1% and 0.027 with errors close to zero values. The highest hit rate of the prediction interval could reach more than 94%, and the trend of FeO composition prediction was consistent with the real situation. This result contributed to achieve accurate advance control of FeO composition trends and values.