How Vetta has been developing digital tools for emission control
By Paula Pomaro, Kássio Cançado, Alexsander Costa, Ana Carolina Rocha, and Lis Soares
In the context of steel mills, nitrogen oxide (NOx) emissions pose a significant challenge due to their polluting properties and the complexity of their formation. Typically, NOx emissions can either be calculated using emission factors multiplied by the amount of fuel burned or measured through Continuous Emission Monitoring Systems (CEMS). Both methods have limitations. Calculations based on emission factors do not account for the complexity of NOx formation, which is largely influenced by temperature. On the other hand, CEMS presents high implementation and operational costs and involves rigorous calibration procedures.
In this scenario, Vetta has been developing machine learning models capable of continuously calculating NOx generation with greater accuracy than emission factor-based calculations and at a much lower cost than CEMS. In a recent case study, Vetta explored the application of machine learning models to predict NOx generation in the reheating furnace of an integrated steel mill.
In the case study, the use of static machine learning models was proposed to predict NOx generation in the furnace stack. The model was trained using NOx data obtained from a CEMS installed in the furnace. The objective was to verify if a machine learning model could predict NOx generation as reliably as the values obtained from a calibrated CEMS. If the machine learning model proved reliable, it would be possible to replace the installation of a CEMS in another furnace with an isokinetic measurement campaign, which would generate the data necessary to train the model. The methodology involved defining input variables, classifying the variables, modeling, and testing. Given that the data followed a non-normal probability distribution, Spearman's Rank Correlation Coefficient was used to identify the variables most correlated with NOx output. The chosen model was Random Forest, whose hyperparameters were optimized for the specific application.
Although the study is still ongoing, it has become evident that there is great potential in the application of artificial intelligence tools for predicting emissions such as NOx. However, certain conditions are necessary:
• Proper process control: To ensure accurate NOx prediction through machine learning models, the combustion system must be well controlled. In the case of furnaces, temperature measurement and control must be rigorous.
• Careful planning of data acquisition campaigns: To ensure the model can adequately predict NOx generation, training data must be obtained under widely varied operating conditions. Therefore, any isokinetic measurement campaign must be carefully planned.
• Customization and adaptation of models: Different processes may require different types of machine learning models and different parameterizations. Therefore, there is no single, off-the-shelf solution. Each process must be fully customized to reflect reality as much as possible.
At Vetta, we are committed to developing solutions that fit perfectly with each client's process reality. We offer engineering solutions with digital tools and continuously develop new functionalities to optimize real-time monitoring, predict operational outcomes, and reduce costs, always aligned with our clients' performance and sustainability goals. Contact us to learn how our solutions can contribute to your business.