Comparison of support vector regression- and neural network-based soft sensors for cement plant exhaust gas composition

Abstract

Application of accurate gas sampler and analyzers is the most reliable method for measuring cement kiln stack gas composition; however, both are very prone to frequent maintenance due to the presence of very fine dust particles in kiln stack gas. Reliable soft sensors are necessary for online concentration estimation of required components during down time of sampler or analyzer. Since accurate first principal models for cement kiln are very complex and time consuming to solve, artificial neural networks and support vector regression are applied as modeling tools. These tools are used to develop four soft sensors for prediction of O2, CO, NO and CH4 in kiln stack gas. A data set consisting of 29,600 data points on 25 process variables collected during a period of 7 month is used for soft sensor development. To have a meaningful comparison on performance, the same methods for plant data processing, feature variable selection and model optimization are applied for both artificial neural networks- and support vector regression-based soft sensors. Refined data set after data processing is divided into training, test and validation groups that contain 70%, 10% and 20% of data set, respectively. Both average absolute error and average absolute relative error calculated for soft sensor predictions revealed that support vector regression-based soft sensors are more accurate compared to corresponding artificial neural networks-based soft sensors.