Research On Anode Effect Prediction Of Aluminum Electrolysis Based On Instance Transfer Learning

The anode effect in the production process of aluminum electrolysis will lead to increased energy consumption and reduced current efficiency,which will result in a reduction in production efficiency.Therefore,timely and effective prediction of the anode effect can help to avoid the above problems.In actual production,due to many problems,the amount of labeled data is insufficient,which greatly limits the performance of anode effect prediction methods driven by traditional data.Recently,transfer learning theory utilizes sufficient labeled data in the source domain to help target domains lacking labeled data to build reliable models,and provides an effective means to solve the above problems.Therefore,this thesis takes the actual production data of two aluminum electrolysis plants in China as research object,and aims at the problem that it is difficult to train a reliable anode effect prediction model with a small amount of labeled data in the target domain,then utilizes transfer learning theory to systematically carry out anode effect prediction research from three aspects: aluminum electrolysis low-quality data preprocessing,source domain samples selection and migration,and aluminum electrolysis data classification.The main work of this thesis is as follows:Firstly,in order to reduce the error in the measurement process,this thesis uses the K-Means algorithm to remove outliers on the original aluminum electrolysis data,then uses the random forest algorithm to fill in the missing values and uses the minimum and maximum standardization method to eliminate the difference of different feature dimensions which provides real data support for subsequent migration.On this basis,in order to solve the problem that the samples in the source domain dataset that are not related to the target domain dataset are prone to negative migration,this thesis carried out a source domain samples screening study based on the K-Means algorithm.By analyzing the distribution of the source domain and the target domain,the source domain samples unrelated to the target domain are removed,then through the improved Tr Adaboost algorithm conduct instance migration.Finally,in order to solve the problem of low efficiency of hyperparameter optimization in the existing anode effect model,this thesis studies and designs an anode effect prediction model based on the Bayesian Optimization(BO)and Light gradient boosting machine(Light GBM)algorithm,which reduces the search range and greatly improves the optimal hyperparameter search efficiency.In this thesis,the validity of the proposed method is verified through experimental analysis.The results show that the F1 score of the anode effect prediction can reach 89.6% when it is 10 minutes ahead of schedule;It is 6.3% higher than the training using only the target domain samples,which realizes the effective prediction of anode effect under small samples condition.