doi:10.16597/j.cnki.issn.1002-154x.2025.05.003
基于 XGBoost 算法的电厂脱硫系统 SO₂ 排放浓度识别
周 富
（国能浙江宁海发电有限公司，宁波 315612）
摘 要 为了有效捕捉电厂脱硫系统 SO₂ 排放特征数据中的复杂关系，提高脱硫出口 SO₂ 浓度识别准确性，研究基于极端梯度提升(XGBoost)算法的电厂脱硫系统 SO₂ 排放浓度识别方法。分析燃煤锅炉燃烧过程以及脱硫系统湿法烟气净化工艺流程，选取影响 SO₂ 排放浓度的特征变量，运用随机森林算法，通过计算特征变量的重要度，筛选出关键特征变量。将关键特征变量输入到基于 XGBoost 的识别模型中，平衡模型预测性能和模型复杂度，确定每个决策树叶子节点的最优权值及最优目标值后，实现电厂脱硫系统 SO₂ 排放浓度识别。实验结果表明：该方法可筛选出影响脱硫系统 SO₂ 排放浓度的 8 个关键特征变量，可实现 SO₂ 排放浓度的准确识别，R² 相对均方误差指标值分别为 0.925,30.2%。
关键词 XGBoost 算法 SO₂ 浓度识别 烟气净化 随机森林 重要度 决策树
Identification of SO₂ Emission Concentration in Power Plant Desulfurization System Based on XGBoost Algorithm
ZHOU Fu
(CHN Energy Zhejiang Ninghai Power Generation Co., Ltd., Ningbo 315612, China)
Abstract In order to effectively capture the complex relationships in the SO₂ emission characteristic data of power plant desulfurization systems and improve the accuracy of SO₂ concentration identification at desulfurization outlets, a method for identifying SO₂ emission concentration in power plant desulfurization systems based on XGBoost algorithm is investigated. The combustion process of coal-fired boilers and the wet flue gas cleaning process of desulfurization systems are analyzed, and characteristic variables affecting the SO₂ emission concentration are selected. Use the random forest algorithm to calculate the importance of the characteristic variables and select the key characteristic variables. By inputting the key characteristic variables into the XGBoost-based recognition measurement model, balancing the prediction performance and complexity of the model, and determining the optimal weights and target values for each decision tree leaf node, the identification of SO₂ emission concentration in the power plant desulfurization system is achieved. The experimental results show that this method can screen out 8 key characteristic variables that affect the SO₂ emission concentration of the desulfurization system. This method can accurately identify the concentration of SO₂ emissions, with R² and RMSE index values of 0.925 and 30.2%, respectively.
Keywords XGBoost algorithm SO₂ concentration identification smoke purification random forest importance level decision tree