PEMS (Predictive Emissions Monitoring System) is arguably the best option to meet the regulatory requirements of CEM (Continuous Emissions Monitoring). This article answers the question about the types of PEMS and which is the best out of the three main types i.e. First Principle, Neural Network and Statistical (Statistical Hybrid).
The regulatory requirement of CEM (Continuous Emissions Monitoring) has been around for almost 40 years and the technology of PEMS (Predictive Emissions Monitoring System) has been around for over 30 years but there still is confusion when it comes to choosing the appropriate type of CEM for a certain site or operating unit.
We as system integrators have dealt with all common types of CEMS (Continuous Emissions Monitoring System). Based on our interaction with clients, we felt the need to present all relevant information related to CEM (Continuous Emissions Monitoring) in one place and give an unbiased picture.
Just like everything else in life, there is no one best, it all depends on your requirements. Take the common example of a car, there is no single best car in the world. You want luxury, you can choose from Audi, BMW, Mercedes etc, you want a good family car you can choose from Honda, Toyota, Nissan etc., you want exotic sports cars, you can choose from Ferrari, Lamborghini, Bugatti etc. All these cars are made to cater to a different clientele with different needs.
Similarly, there are different types of CEM, the traditional hardware based and software based called PEMS, all made for different requirements. In this article, we will only discuss the different types of PEMS and which is the best for your site & requirements. Discussion about different types of traditional physical gas analysers for CEM will be done in another article.
Before we get into the main topic, it would be worth everyone’s while to have a primer on CEMS (Continuous Emissions Monitoring System).
Combustion processes in Reformers, Boilers & Gas Turbines produce several by-products e.g. NOx, CO, CO2 etc. These by-products can cause serious damage to the environment through smog, acid rain etc. To control & reduce this damage, the US EPA (United States Environmental Protection Agency) has mandated to measure & monitor these emissions. Some of these emissions are required to be monitored on a continuous basis. This is done by CEMS (Continuous Emissions Monitoring Systems).
There are two main type of CEMS:
- Hardware-based CEMS
- Software-based CEMS
Hardware-based CEMS is commonly known just as CEMS, whereas a Software-based CEMS is known as PEMS (Predictive Emissions Monitoring System), so:
- CEMS: Is hardware-based and uses physical gas analyzers to measure the emissions.
- PEMS: Is software-based and uses mathematical algorithms to “predict” the emissions.
History of PEMS
- PEMS was developed in late 1970s.
- Installed first time & approved by US EPA in 1993.
- Since then installed at hundreds of locations worldwide.
- Both CEMS & PEMS are approved by US EPA & other regulatory authorities around the world.
There are three main types of PEMS:
- First principle type
- Neural network type
- Statistical type
First Principle:
The PEMS which work on the First Principal method rely on mathematical correlations based on the fundamentals of emission formation. Emissions from combustion processes can be divided into two classes:
- Stoichiometric Emissions – dictated alone by the composition of the fuel e.g. SO2 & CO2
- Rate Formation Emissions – where emission formation is depending on the conditions (temperature, pressure etc.) in the combustion process e.g. NOx and CO.
As compared to other types of PEMS, the First Principle models require more modeling work. Each model needs to be configured according to the specific combustion source configuration e.g. Boilers, Reformers, Gas Turbines etc. and input signals availability.
Over the years, First Principle PEMS has developed an extensive library of models for specific combustion sources and for different collection of available input signals.
The major advantage of the first principle models is that they exploit the physical facts and use their relations in the model. There are less calibration measurements required compared with other PEMS types.
The disadvantage of the First Principle model is that it can only be developed by someone who has deep understanding of the combustion process for that specific equipment.
Neural Network:
The PEMS which work on the neural network principle develop a relation between inputs and outputs that are generated through a number neuron layers. In each neuron layer there are number of neurons that relate to each neuron in the previous layer and the next layer. Each connection is having a weight that is used for the network calculation.
The fundamental idea in a neural network is to find the right weights for the connections. This is done by “training” the network with input values and known output values. To be accurate the neural network must be trained over the complete operating envelope (combination of input signals).
The important part of setting up a neural network is to define which input signals should be included in the network. Especially parameters with small variation during training can induce error in the model. For that reason, it is necessary to train the network under different operating conditions, different load levels, different fuel composition etc. to have an accurate PEMS system.
In Neural Network systems, those which rely on Modular Neural Network and use RPA (Robotic Process Automation) are superior to simple Artificial Neural Networks.
The major advantage of data driven models is that they apparently do not need any knowledge of the system they are applied on.
On the downside is the need for extensive model training. To have a decent data driven PEMS system that works under all conditions it should have calibration measurements done under various operation condition including changing ambient conditions.
Statistical Model:
As the name suggests, this type of PEMS is based on using some form of statistical modeling e.g. regression modeling, wherein an empirical correlation is developed between various process inputs and emissions rates.
The statistical model can be developed for any given process without any knowledge of the underlying combustion chemistry or process knowledge. To develop this model, the emissions generating unit is run through its entire operating range for a limited time and all the primary, secondary and tertiary process inputs are logged. At the same time, the generated emissions which need to be monitored are logged.
Various statistical modeling techniques are then used to develop a correlational model which represents the actual process.
The downside of this model is that although it doesn’t require the knowledge of the process or chemistry, it requires deep knowledge of statistical modeling.
RM (Reference Method):
Once the model is developed, it is tested against a RM for accuracy. The US EPA PS-16 governs this certification & testing specifically for the PEMS. In addition to PS-16, US CFR60, CFR75 and their relevant subparts govern the testing, accuracy, auditing of the PEMS.
CONCLUSION:
Here are the key takeaways:
- If your process has changing fuel properties e.g. an incinerator where the feed composition varies hugely, then you will be better off with a physical analyser CEMS.
- If your process has fairly constant fuel composition e.g. natural gas or fuel oil, then you can safely install a PEMS.
- All three types of PEMS meet regulatory requirements for accuracy & reliability.
- The Neural Network PEMS are usually more expensive than the First Principle or Statistical PEMS but can arguably offer more accuracy also.
- The First Principle & Statistical models have almost similar accuracy & cost.
- The First Principle models require complete understanding of the combustion process & chemistry and if your process ever changes in future, you will need the PEMS vendor to update the model.
- The Statistical model doesn’t require understanding of the process and if your process ever changes in the future, you can update the model yourself with some training from the vendor.
So in the end it depends on your requirements and how your organization operates, if most of the work is done in-house and you want to develop in-house resources, then you should look at the statistical model favorably but if you contract out most of the work like most organizations, then Neural Networks should be your first choice followed by First Principle.
Please Contact Us in case you require just consultation or even complete installation of PEMS at your site.
You can read further about PEMS and how our PEMS solutions can help you by following this link https://www.technology-edge.com/pems-predictive-emissions-monitoring-system/
Please contact us if you would like to explore whether PEMS is a good option for you:
