An adaptive least squares support vector machine model with a novel update for NOx emission prediction

a r t i c l e i n f o Keywords: Data-driven model Model update Least squares support vector machine NOx emissions Coal-fired boiler This paper presents an adaptive least squares support vector machine (LSSVM) model with a novel update to tackle process variations. The key idea of the update is to divide the process variations into two main categories, namely, irreversible and reversible variations. Correspondingly, sample addition and sample replacement are proposed to update the model. The incremental LSSVM algorithm and detailed update procedure are also provided. A benchmark simulation with a time-varying nonlinear function is conducted to evaluate the effectiveness of the update algorithm. Finally, the proposed method is applied to predict the nitrogen oxide (NOx) emissions of a coal-fired boiler using real operation data from a power plant. Results reveal that the LSSVM model with the novel update maintains high prediction accuracy despite different process characteristics. Meanwhile, the time consumed in the update process is decreased because of the incremental form compared with the model reconstruction. In industrial processes, certain types of primary variables in industrial processes, such as product qualities and flue gas concentration, should be measured accurately and reliably at all times. These variables may be required to be maintained within specified limitations in accordance with government regulation and manufacturing criteria, besides , they are also very important in guiding optimal operation [1,2]. However, the accurate measurement of primary variables is hindered by high costs and technical limitations. Although online analyzers are available in a number of plants, these hardware-based instruments are not only highly vulnerable to failure because of being operated under harsh environments but are also expensive and difficult to maintain. Therefore, a new way of realizing a redundant measurement has crucial significance to ensure the safe, economical and efficient operation of plants [3,4]. Soft sensor techniques are widely accepted for estimating primary variable with the use of other relevant variables that are easy to measure online. Soft sensors can operate in parallel with hardware-based sensors, thereby providing a backup and redundant measurement. Moreover, if the soft sensor model that describes the relationship between the primary and other operating variables is developed, the process operation can also be optimized by regulating the operating parameters [5,6]. Soft sensor models are generally established on the basis of first-principle and data-driven methods. Although often desirable, first-principle models are impractical in most cases because they involve …