Pulverized coal injection (PCI) is a dominant technology in ironmaking blast furnaces (BFs) for energy efficiency and cost reduction, while the relevant in-furnace phenomena are experimentally inaccessible. It is desired to understand these in-furnace phenomena in a timely manner. In this study, a data-driven approach is developed for rapid predicting the multi-objective in-furnace combustion characteristics related to PCI operation in a BF. The approach includes a database of computational fluid dynamics (CFD) 243 simulations in terms of flow field, temperature field, gas species concentration and coal burnout within the raceway; and a machine learning (ML) model where random forest regression model is selected due to its higher accuracy than others. The results show that this approach can predict the multi-objective in-furnace phenomena with high accuracy in aspects of temperature, gas species concentrations and combustion efficiency in the raceway. Furthermore, three additional cases - no. 244-246 scenarios outside the database, is tested to demonstrate the ML prediction effectiveness through virtualizing and comparing the full in-furnace phenomena. The response time of this approach is nearly 16,000 times shorter than the CFD simulations while achieving similar accuracy. This prediction approach provides a time- and cost-effective tool for optimizing the responses of in-furnace phenomena to PCI operation changes.

喷煤 (PCI) 是炼铁高炉 (BFs) 中用于提高能源效率和降低成本的主要技术，而相关的炉内现象在实验上是无法实现的。希望及时了解这些炉内现象。在这项研究中，开发了一种数据驱动的方法，用于快速预测与高炉中 PCI 操作相关的多目标炉内燃烧特性。该方法包括计算流体动力学 (CFD) 数据库 243 模拟流场、温度场、气体物质浓度和滚道内的煤燃尽；和机器学习 (ML) 模型，其中随机森林回归模型因其比其他模型更高的准确性而被选中。结果表明，该方法可以从温度、气体种类浓度和滚道燃烧效率等方面对多目标炉内现象进行高精度预测。此外，另外三个案例 - 没有。对数据库外的 244-246 个场景进行测试，通过虚拟化和比较完整的炉内现象来证明 ML 预测的有效性。这种方法的响应时间比 CFD 模拟短近 16,000 倍，同时实现了相似的精度。这种预测方法为优化炉内现象对 PCI 操作变化的响应提供了一种节省时间和成本的工具。对数据库外的 244-246 个场景进行测试，通过虚拟化和比较完整的炉内现象来证明 ML 预测的有效性。这种方法的响应时间比 CFD 模拟短近 16,000 倍，同时实现了相似的精度。这种预测方法为优化炉内现象对 PCI 操作变化的响应提供了一种节省时间和成本的工具。对数据库外的 244-246 个场景进行测试，通过虚拟化和比较完整的炉内现象来证明 ML 预测的有效性。这种方法的响应时间比 CFD 模拟短近 16,000 倍，同时实现了相似的精度。这种预测方法为优化炉内现象对 PCI 操作变化的响应提供了一种节省时间和成本的工具。