Novel method for temperature prediction in rotary kiln process through machine learning and CFD
Yaozu Wang,Yue Xu,Xiaoran Song,Qingke Sun,Jianliang Zhang,Zhengjian Liu
DOI: https://doi.org/10.1016/j.powtec.2024.119649
IF: 5.64
2024-03-16
Powder Technology
Abstract:Blast furnace high-proportion pellet smelting stands as a pivotal pathway towards achieving low-carbon and environmentally friendly ironmaking. However, challenges in pellet ore production, particularly through the rotary kiln method, have surfaced due to the scarcity of iron ore resources, leading to issues like "ringing" that hamper production efficiency and energy conservation. Effectively predicting and optimally controlling the temperature field within rotary kilns emerges as a crucial strategy to address this pressing challenge. In a pioneering effort, this paper introduces a data-driven machine learning methodology designed to expedite the prediction of temperature field distributions within rotary kilns. Through a detailed analysis of field data from the factory and the insights of experts, the paper has discussed on the key factors influencing the temperature field dynamics in rotary kilns. Leveraging Computational Fluid Dynamics (CFD), a numerical simulation model was established to replicate rotary kiln oxy-coal combustion scenarios. A comprehensive dataset was generated using high-throughput computing technology, encompassing a variety of conditions such as coal injection rates, return air temperature, return airflow volume, fixed carbon content of pulverized coal, and volatile fraction of pulverized coal. This dataset included a total of 625 distinct cases. Subsequently, employing machine learning techniques including Random Forest (RF) and Genetic Algorithm-optimized Random Forest (GA-RF), predictive models were developed for rotary kiln temperature field dynamics. A rigorous comparative analysis was conducted to explore the impact of different machine learning dataset sizes on the predictive accuracy of the models. The results of this analysis demonstrate that, enhanced by the genetic algorithm of RF, it surpasses the performance of the standard RF approach. This enhancement resulted in a significant improvement in prediction accuracy, as evidenced by the reduction of the Root Mean Square Error (RMSE) from 25.0879 to 17.5330 and the Mean Absolute Error (MAE) from an initial value of 6.5600 to 4.7661. Furthermore, the novel soft-measurement method introduced in this paper significantly reduces simulation time by several orders of magnitude. It has diminished the response time from the traditional industrial methods, which took two to three weeks, to approximately 10 s, as compared to conventional CFD approaches. Remarkably, this approach successfully preserves a high level of accuracy. The implications of these research findings are far-reaching,providing novel insights into the acceleration of CFD simulation calculations and the design of soft measurement sensors customized for black-box reactors in industrial processes.
engineering, chemical