Achieving optimization and control of industrial processes relies on accurate models. This paper proposes a combined data model for modeling complex industrial processes, utilizing a graph convolutional network and an improved Transformer model. The process variables’ correlations are firstly determined based on domain knowledge, followed by aggregating neighboring node information to enhance node feature representations via graph convolutional network. The multi-head self-attention mechanism of the Transformer model is then utilized to extract global features and finally map them to process indicators through the fully connected layer. This paper introduces the concept of graph mask matrix to improve the classical Transformer model. The graph mask matrix suppresses the generation of interfering information by zeroing attention weights among irrelevant variables, enhancing the Transformer model’s modeling capability model. The effectiveness of the proposed strategies is verified with extensive simulation experiments on natural industrial objects.