Research And Application Of Prediction Algorithm For Resistivity Of Anode Calcined Coke Powder

China is an important production base of prebaked anode in the world.The resistivity of calcined coke powder affects the quality of prebaked anode to a certain extent.It is of great significance to study the influencing factors of the resistivity of calcined coke powder and predict the resistivity by using machine learning algorithm,which is convenient for the technicians to adjust the production control parameters in time,and to improve the quality of anode.Firstly,according to the principle of anode production process,the historical data of anode calcined coke production are preprocessed and visualized.Secondly,the material parameters and control parameters that affect the resistivity of calcined coke powder are searched,and the method of inverting the predicted value using XGBoost is explored to assist the feature screening.Aiming at the problems of low accuracy and slow training speed of GABP algorithm,the Xavier initialization is introduced to replace the initialization process of BP randomly generating weight and threshold;Aiming at the problem that the algorithm is prone to local convergence,the LM(Levenberg-Marquarelt)algorithm combining gradient descent and Gauss-Newton method is introduced,which effectively improves the situation that the model training tends to fall into local convergence,and the algorithm training process is implemented in practice.For optimization,a new XLG-BP algorithm is constructed by combining these techniques.The cubic function fitting experiments show that the XLG-BP algorithm is superior to the traditional and improved algorithms in performance indicators.Based on the research of algorithm,the XLG-BP algorithm is applied to the prediction of the resistivity of calcined coke powder,and a system for predicting the resistivity of calcined coke powder for anode is developed.The functions of visual analysis and preprocessing of anode calcination production data,prediction algorithm of calcined coke powder resistivity,model training,operation,evaluation and visual display of results are realized.