Neural networks for boiler emission prediction

In the USA and throughout the world, Pulp and Paper mills face increasingly stringent regulations, regarding air emissions from their power boilers. Mill managers must select the proper methods and technologies to meet environmental regulations. The most common technique for measuring emissions is a combination of hardware analyzers known as continuous emissions monitoring systems (GEMS). Predictive emissions monitoring systems (PEMS) are an alternative air monitoring method now gaining wide acceptance in the USA. These systems use neural network modeling techniques that are applied in powerful software engines which infer emission measurements through relationships with other known process limits. PEMS therefore can determine stack contaminants in real time by correlating emissions with key combustion unit parameters like fuel type, air and fuel flows, combustion temperatures, etc. Most of this data is available from the process control system, but for training the net there must also be simultaneous collection of the stack emissions information that can be provided with portable monitoring equipment. Predictive emissions monitoring systems based on neural networks can now be considered a viable way to combine accurate and reliable monitoring, with low purchase, installation, and maintenance costs. The paper discusses steps required to implement a neural network model for this application, as well as the issues to consider in verifying the final output from the model. Furthermore, a comparison of using a neural net PEMS system versus the traditional GEMS technology for emission monitoring is presented.