Predictive emissions monitoring systems (PEMSs) are alternatives to continuous emissions monitoring systems (CEMSs) for monitoring air pollutants, such as NO${}_{\mathrm{x}}$. Existing PEMSs and related research have focused on applying artificial neural network (ANN) algorithms. However, ANN-based models are treated as “black boxes”. Regulators and decision makers without a statistical background often have difficulty understanding these models, which poses a significant challenge for a broader application of PEMSs. In this study, we proposed a tree-based ensemble method with gradient boosting techniques for PEMS development. Compared to ANNs, tree-based methods are easier to understand and require less effort to preprocess data, fewer hyperparameters for model tuning, and less time for model training. We developed a predictive model using a gradient boosting machine learning library called XGBoost to monitor NO${}_{\mathrm{x}}$ emissions from a boiler located in Alberta, Canada. The model uses five process parameters as inputs and the predicted NO${}_{\mathrm{x}}$ emissions as output. We trained the model with 202,047 samples using random search methods to determine the best model and tested the model with 50,512 samples. We evaluated the test results against US EPA PEMS standards. The model passed all the statistical tests for precision outlined by US EPA Performance Specification 16. The Pearson correlation r value was 0.98 between the XGBoost-predicted NO${}_{\mathrm{x}}$ values and the CEMS-measured NO${}_{\mathrm{x}}$ values. The RMSE was 0.14, and the MAE was 0.09. We conclude that XGBoost is a good option for developing PEMSs. Facility operators can use the method provided in this study to develop PEMSs by themselves using the open-source library XGBoost at no cost.

预测性排放物监测系统（PEMS）是连续排放物监测系统（CEMS）的替代品，用于监测空气污染物，例如NO${}_{\mathrm{X}}$。现有的PEMS和相关研究都集中在应用人工神经网络（ANN）算法上。但是，基于ANN的模型被视为“黑匣子”。没有统计背景的监管者和决策者通常很难理解这些模型，这对于更广泛地使用PEMS构成了重大挑战。在这项研究中，我们提出了一种基于树的集成方法，并采用梯度增强技术进行PEMS开发。与人工神经网络相比，基于树的方法更易于理解，并且预处理数据所需的工作更少，用于模型调整的超参数更少，并且用于模型训练的时间更少。我们使用称为XGBoost的梯度提升机器学习库开发了一种预测模型，以监控NO${}_{\mathrm{X}}$位于加拿大艾伯塔省的锅炉产生的废气排放。该模型使用五个过程参数作为输入和预测的NO${}_{\mathrm{X}}$排放作为产出。我们使用随机搜索方法对202,047个样本进行了训练，以确定最佳模型，并使用50,512个样本对模型进行了测试。我们根据美国EPA PEMS标准评估了测试结果。该模型通过了所有统计测试，精确度达到了US EPA性能规范16的要求。XGBoost预测的NO之间的Pearson相关r值为0.98。${}_{\mathrm{X}}$ 值和CEMS测得的NO${}_{\mathrm{X}}$价值观。RMSE为0.14，MAE为0.09。我们得出结论，XGBoost是开发PEMS的不错选择。设施运营商可以免费使用开源库XGBoost自行使用本研究中提供的方法来开发PEMS。