| With the rapid development of digitalization in the aluminum industry,an increasing amount of data is being generated during the electrolytic aluminum production process.The production data contains information such as electrolytic reaction status and process parameter changes,which can assist in precise control of the production equipment in the electrolytic aluminum production process.Currently,there are issues in the processing of electrolytic aluminum production parameters,including fuzzy relationships between process variables,difficulties in tracing electrolytic faults,low accuracy in predicting aluminum output,and insufficient explanatory power for factors affecting aluminum output.To address these issues,a series of temporal processing techniques are employed,and the main research contents are as follows:Firstly,to solve the problem of fuzzy relationships between variables,the Pearson correlation coefficient method combined with normality test is used to obtain the correlation coefficients between variables.Additionally,the Granger causality test combined with stationarity test is used to determine the causal relationships between input and output sequences.A directed probabilistic graphical model is utilized to represent the causal relationships and correlation coefficients between process variables.This model integrates the dependencies between variables,providing an accurate and intuitive representation of variable relationships.It has been widely promoted and applied in the analysis department of aluminum plants.Secondly,to address the issue of difficulties in fault tracing,a fault tracing model is established by incorporating fault phenomena and causes into the directed probabilistic graphical model.This model associates faults with their causes and with other faults,enabling quick identification of fault causes and prevention of secondary faults.Thirdly,to improve the accuracy of aluminum output prediction,the MLP neural network,suitable for multi-dimensional time series data,is employed to train and predict aluminum output factors,significantly improving the prediction accuracy.To further enhance the model’s performance,features such as time series data embedding and concatenation,convolutional modules,and attention mechanisms are incorporated into the MLP model.Feature concatenation and convolution can extract deep features from time series data,while the attention mechanism enhances the model’s focus on important features.Experimental results demonstrate the effectiveness of the model improvements.Finally,to address the insufficient explanatory power of factors affecting aluminum output,a multi-level regression model is established by combining the autoregressive model and the multivariate regression model for aluminum output.The weights of the two models are obtained using Bayesian estimation.This complete aluminum output prediction mathematical model considers the influences of historical aluminum output and current other variables on the current aluminum output,with coefficients of 0.072 and 1.146,respectively.This model effectively explains the input,output,and intermediate calculation processes of aluminum output.Experimental validation confirms the effectiveness of the multi-level regression model. |
