In the aluminum smelting process, real-time measurement of the furnace temperature is a key factor to ensure the production of aluminum. However, in a complex furnace environment, it is sometimes difficult or costly to measure the temperature accurately. To address this issue, a soft sensor furnace temperature prediction model is proposed. The sensor model utilizes a graph convolutional network (GCN) to extract the spatial features of the variables and a long short-term memory network (LSTM) to extract the temporal features of the variables. Firstly, a GCN with a correlation aggregation strategy (CAGCN) is proposed to reasonably aggregate neighborhood node features and filter out less important features. Secondly, to efficiently allocate attention weights, an LSTM with a dynamic attention mechanism (DALSTM) is developed. Finally, CAGCN and DALSTM are synergistically used to predict the furnace temperature in the aluminum smelting process. The performance of the model is verified by five indices such as the coefficient of decision (<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="13042_2025_2651_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">$R^2$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>) and the maximum absolute error (MAX). The <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="13042_2025_2651_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">$R^2$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> of the proposed model on the two testing sets is 0.99820 and 0.99632, respectively, and the MAX is 9.5883 and 10.3668, respectively. The results of the experiments show that the proposed model meets the accuracy requirements of the aluminum smelting process.

Aluminum smelting temperature prediction using GCN with correlation aggregation strategy and LSTM with dynamic attention

期刊论文
发表日期：

2025年4月22日

查看全文

Jiayang Dai,
Haofan Shi,
Xingyu Chen,
Hangbin Liu,
Peirun Ling

摘要

In the aluminum smelting process, real-time measurement of the furnace temperature is a key factor to ensure the production of aluminum. However, in a complex furnace environment, it is sometimes difficult or costly to measure the temperature accurately. To address this issue, a soft sensor furnace temperature prediction model is proposed. The sensor model utilizes a graph convolutional network (GCN) to extract the spatial features of the variables and a long short-term memory network (LSTM) to extract the temporal features of the variables. Firstly, a GCN with a correlation aggregation strategy (CAGCN) is proposed to reasonably aggregate neighborhood node features and filter out less important features. Secondly, to efficiently allocate attention weights, an LSTM with a dynamic attention mechanism (DALSTM) is developed. Finally, CAGCN and DALSTM are synergistically used to predict the furnace temperature in the aluminum smelting process. The performance of the model is verified by five indices such as the coefficient of decision ( $R^2$ $R^{2}$ ) and the maximum absolute error (MAX). The $R^2$ $R^{2}$ of the proposed model on the two testing sets is 0.99820 and 0.99632, respectively, and the MAX is 9.5883 and 10.3668, respectively. The results of the experiments show that the proposed model meets the accuracy requirements of the aluminum smelting process.