Research On Prediction Of Aluminum Electrolyzer Condition Based On TCN

Aluminum electrolysis is an industrial process system for the production of aluminum.Tank condition refers to the production capacity of the aluminum electrolyzer during production.A good tank condition helps to improve aluminum electrolysis production efficiency and extend the life of the cell.During the aluminum electrolysis production process,production personnel collect a large amount of historical production data.These data are characterized by multi-dimensional time series and contain a large amount of production information.In this paper,we propose a method based on temporal convolutional networks using deep learning methods to analyze and evaluate the electrolytic cell conditions,so that production personnel can adjust the cells in time and improve production efficiency.The main content of the research in this paper contains the following four aspects.(1)An improved whale optimization algorithm is proposed.To address the problem that random initialization may lead to insufficient optimization of initial population positions,we propose to integrate the use of tent mapping initialization method in the whale optimization algorithm to generate population positions with randomness and fractal characteristics to increase the diversity of search and the ability of local search.To address the problem that the linearly decreasing convergence factor of the whale optimization algorithm may lead to premature convergence of the algorithm in some cases,a nonlinear convergence factor is introduced to balance the global search ability and local search ability.To address the problem that the whale optimization algorithm tends to fall into local optimal solutions prematurely,inertia weights are introduced to keep each individual in the previous direction during the iterative process,so that individuals are influenced by the combination of historical optimal solutions and current individual positions to better achieve diverse search and fast convergence.(2)The IWOA-MMTCN-AT network is proposed.In the original TCN network,the structure of the temporal convolutional network is one-dimensional convolution,which leads to the problem that in the multidimensional timing problem,there is only sequential input in the one-dimensional convolution,which eventually leads to poor prediction results.In this topic,for multidimensional aluminum electrolyzer data,the convolutional structures of several TCNs are connected in parallel for the input of multidimensional data to prevent crosstalk between different data dimensions and improve the prediction performance of the model.Meanwhile,to address the problem that the fully connected layer in the temporal convolutional network will ignore the correlation between different positions in the sequence,an attention mechanism is introduced to assign different weights according to the correlation in the temporal data to accurately capture the correlation between the data.The Leaky Re LU function is also introduced to make the TCN network deliver the gradient signal faster and alleviate the gradient disappearance problem.For the problem that hyperparameters need to be selected manually,the improved whale optimization algorithm proposed in this paper is selected to perform hyperparameter search for the temporal convolutional network and improve the accuracy of the algorithm.(3)A TCN-based aluminum electrolyzer condition prediction model was established.Experiments were conducted using historical data of aluminum electrolytic cells.The experiments show that the model outperforms the LSTM model and the standard TCN model in terms of performance.(4)An aluminum electrolyzer condition prediction and analysis system was designed and implemented.The system implements the function of storing and viewing the aluminum electrolysis data,the aluminum electrolysis tank prediction algorithm proposed in this paper,and can display the production data and tank condition data of aluminum electrolysis tank in various ways.