Development of Predictive Emissions Monitoring System Using Open Source Machine Learning Library – Keras: A Case Study on a Cogeneration Unit

Minxing Si,Brett M Wiens,Tyler J Tarnoczi,Ke Du

doi:10.1109/access.2019.2930555

Abstract

The study provides an overview of Predictive Emissions Monitoring System's (PEMS) research, application, installation, and regulatory framework as well as develops predictive models for NOx emissions from a natural gas fired cogeneration unit using an open source machine learning library, Keras, and open source programming languages, Python and R. Nine neural network based predictive models were trained with 12 086 examples and tested with 3020 examples. The neural network-based models use eight process parameters as inputs to predict NOx emissions. All models meet the regulatory requirements for precision. The best model (32-64-64-64) has four hidden layers and uses the Nadam method for optimization. The best model has a mean absolute error of 0.5982, r-value of 0.9451, and a difference of 0.14% between the measured and predicted emission values using the test dataset. The study demonstrated the feasibility of using open source machine learning library in PEMS development. It also provides guidance to facility operators to develop their own PEMS models for monitoring emissions.

Highlights

The Predictive Emissions Monitoring System (PEMS) was developed as an alternative to overcome the drawbacks of the Continuous Emissions Monitoring System (CEMS), such as high initial capital cost, high operating cost, maintenance and operator training [1]
Nine models that have the minimum average mean square error (MSE) and mean absolute error (MAE) in the 30 repeats were chosen for statistical tests using the entire training dataset of 12,086 training examples
Nine neural networks based on models with different structures and optimization methods were tested to predict NOx emissions using eight process parameters

Summary

Introduction

The Predictive Emissions Monitoring System (PEMS) was developed as an alternative to overcome the drawbacks of the Continuous Emissions Monitoring System (CEMS), such as high initial capital cost, high operating cost, maintenance and operator training [1]. A PEMS relies on using the operating parameters of combustion facilities through first principle, statistical or Artificial Intelligence (AI) methods to build a model that can predict emissions. The associate editor coordinating the review of this manuscript and approving it for publication was Canbing Li. The capital costs for PEMS are estimated to be 50% less than for CEMS, and the operations and maintenance costs are approximately 10–20% of the CEMS cost [2], [3]. The application of the PEMS includes a) compliance reporting, b) offline what-if analysis, c) analyzer availability enhancement and d) continuous estimating when CEMS is offline [4]

Objectives

Results

Conclusion