Two-dimensional temperature field prediction of rotary kiln based on graph neural networks

Sensing and optimizing the temperature distribution in rotary kilns is key to improve energy efficiency and reduce production costs. Traditional computational fluid dynamics (CFD) solvers are computationally expensive and cannot meet the demand for real-time performance in industrial sites. With the continuous development of deep learning, graph neural networks (GNNs) have emerged as a potentially effective method for accelerating CFD unstructured grid simulations. In order to accurately predict the whole temperature field in a rotary kiln, a novel GNN model is designed in this study, and the CLJPNet model is proposed for fast prediction of the whole temperature field in a rotary kiln. Compared with the traditional GNN, this study is able to accurately predict the rotary kiln temperature field by using the Cleary-Luby-Jones-Plassmann Coarsening coarsening algorithm in the multi-algebraic lattice to sparsify the graph topology to accelerate the inference speed while maintaining a high accuracy. Finally, the model proposed in this paper is compared with the other three models to verify the effectiveness of the model. The experimental results indicate that the model proposed in this study achieves a coefficient of determination (R2) of 0.99, mean squared error of 710.63, mean absolute percentage error of 1.64, and mean relative error in the region of interest of 0.02 on the test set, and all evaluation metrics are superior to other models, demonstrating better prediction performance. In addition, the proposed model runs 3 orders of magnitude faster than the CFD model. The rapid prediction method for temperature fields proposed in this paper provides a novel approach to the intelligent advancement of rotary kiln production.