Production prediction and energy saving of complex industrial processes using multiscale variable dynamic interaction information extraction network integrating regressor

In complex industrial processes, process variables interact with each other, and the extraction of these interaction features is crucial for industrial process modeling. Traditional methods can extract the global correlation between variables, but cannot effectively extract dynamic interaction relationships between variables. Therefore, a novel Multiscale Variable Dynamic Interaction Information Extraction Network (MSVE) integrating Regressor (REG) (MSVE-REG) is proposed to model the interaction characteristics of process variables at different scales. The multiscale decomposition method (MSD) is designed to convert raw industrial process data into two-dimensional tensors based on multiple periods to construct multiscale industrial data. Then, the MSVE constructs multiscale adjacency matrices for the graph attention neural network (GAT) based on multiscale industrial process data to extract dynamic interaction features between variables at different scales. Moreover, the REG establishes the mapping from dynamic multiscale features to the output. Finally, a MSVE-REG based production prediction model is constructed to improve energy conservation and efficiency in the actual gasoline production process. The experimental results show that compared with other prediction models, the MSVE-REG has achieved the most accurate results, with the mean absolute error, the root mean square error, the mean absolute percentage error and the R-square reaching 0.00032, 0.00101, 0.00018 and 0.98521, respectively, providing energy saving of industrial production processes.