Enhanced PEMS Performance and Regulatory Compliance through Machine Learning

Modeling technologies can provide strong support to existing emission management systems, by means of what is known as a Predictive Emission Monitoring System (PEMS). These systems do not measure emissions through any hardware device, but use computer models to predict emission concentrations on the ground of process data (e.g., fuel flow, load) and ambient parameters (e.g., air temperature, relative humidity). They actually represent a relevant application arena for the so-called Inferential Sensor technology which has quickly proved to be invaluable in modern process automation and optimization strategies (Qin et al., 1997; Kadlec et al., 2009). While lots of applications demonstrate that software systems provide accuracy comparable to that of hardware-based Continuous Emission Monitoring Systems (CEMS), virtual analyzers are able to offer additional features and capabilities which are often not properly considered by end-users. Depending on local regulations and constraints, PEMS can be exploited either as primary source of emission monitoring or as a back-up of hardware-based CEMS able to validate analyzers’ readings and extend their service factor. PEMS consistency (and therefore its acceptance from environmental authorities) is directly linked to the accuracy and reliability of each parameter used as input of the models. While environmental authorities are steadily opening to PEMS, it is easy to foresee that major recognition and acceptance will be driven by extending PEMS robustness in front of possible sensor failures. Providing reliable instrument fail-over procedures is the main objective of Sensor Validation (SV) strategies. In this work, the capabilities of a class of machine learning algorithms will be presented, showing the results based on tests performed actual field data gathered at a fluid catalytic cracking unit