Gated Recurrent Knowledge-Guided Attention Network with Adaptive Graph Structure Learning for Forecasting in Chemical Processes

The prediction of key variables in chemical processes presents significant challenges due to the intricate dynamic relationships among the variables. Most existing graph-based methods for chemical processes rely on a static topological graph structure derived from predefined variable relationships and only consider pairwise relationships between variables, which are inconsistent with practical scenarios. To overcome these limitations, this paper proposes the gated recurrent knowledge-guided attention network with adaptive graph structure learning (GRKAT-GSL), an end-to-end recurrent graph neural network for knowledge-guided graph attention calculation based on the learned topological graph structure. First, variables are embedded using different convolution kernels to extract features, which are then utilized to construct a topological graph as the input to the graph recurrent neural network. Then, a graph attention calculation method, the knowledge-guided attention mechanism (KGAM), is proposed to calculate the relationships between connected nodes in the graph to guide the transmission of information between nodes in the recurrent network. Experimental results on the Tennessee Eastman (TE), fluid catalytic cracking (FCC), and METR-LA data sets demonstrate that GRKAT-GSL outperforms other methods. Furthermore, the analysis of the experimental results indicates that the proposed method is capable of capturing complex relationships in chemical processes.