Contact:

Abstract

It is possible to carry out a diagnosis of the operation of a gas turbine based on the measurement of its polluting gases produced by its combustion, especially CO and NOx, together with other operating parameters such as power, fuel gas consumption , ambient temperature, axial compressor discharge pressure and stage nozzle temperatures, analyzing the trend of their values and applying machine learning algorithms to predict their values.

A variation in the time of the mixture between the fuel gas and air inside the combustion cans can be caused by aging and variations in parameters such as the axial compressor discharge pressure , can lead to an increase in emissions.

In the following figures, not related to the data processing below, they are an example of how the combustion temperature and oxygen are correlated with the concentration of NOx and CO

The operating power is an important factor, in the case of carbon monoxide (CO) it takes preponderance at low load, whereas Nitrogen Dioxide (NOx) takes importance at full load with high flame temperatures.

Data Set

We are going to exemplify by applying machine learning algorithms to data obtained from the UCI Machine Learning Repository at the following link:

https://archive.ics.uci.edu/ml/datasets/Gas+Turbine+CO+and+NOx+Emission+Data+Set

Dua, D. and Graff, C. (2019). UCI Machine Learning Repository [http://archive.ics.uci.edu/ml]. Irvine, CA: University of California, School of Information and Computer Science.

The dataset contains 36733 instances of 11 sensor measures aggregated over one hour (by means of average or sum) from a gas turbine located in Turkey’s north western region for the purpose of studying flue gas emissions, namely CO and NOx (NO + NO2). The data comes from the same power plant as the dataset used for predicting hourly net energy yield. By contrast, this data is collected in another data range (01.01.2011 - 31.12.2015), includes gas turbine parameters (such as Turbine Inlet Temperature and Compressor Discharge pressure) in addition to the ambient variables.

Attribute Information:

The explanations of sensor measurements and their brief statistics are given below.

Variable Abbr Unit Min Max Mean
Ambient temperature AT C 6.23 37.10 17.71
Ambient pressure AP mbar 985.85 1036.56 1013.07
Ambient humidity AH % 24.08 100.20 77.87
Air filter difference pressure AFDP mbar 2.09 7.61 3.93
Gas turbine exhaust pressure GTEP mbar 17.70 40.72 25.56
Turbine inlet temperature TIT C 1000.85 1100.89 1081.43
Turbine after temperature TAT C 511.04 550.61 546.16
Compressor discharge pressure CDP mbar 9.85 15.16 12.06
Turbine energy yield TEY MWH 100.02 179.50 133.51
Carbon monoxide CO mg/m3 0.00 44.10 2.37
Nitrogen oxides NOx mg/m3 25.90 119.91 65.29

Axial compressor discharge pressure (CDP) prediction

Taking into account that the flame temperature and the O2 concentration depend on the discharge pressure of the axial compressor, we will subsequently make a prediction of that pressure, applying a “gradient boost” machine learning regression algorithm, trained with values before the 2013.

The features variables to predict CDP are: AT, AP, AH, AFDP, GTEP, TEY.

With the same features, we will make a prediction of the turbine inlet temperature (TIT) as well, to compare values.

Below we show the difference between the real and predicted value of the axial compressor discharge pressure:

We observe that the differences arise from 2014 approximately, perhaps showing a problem related to IGV, por example.

Pre swirl effect

Inside the combustion can, air and fuel gas are mixed, through special injection geometries, in which eddies are created for each fluid, especially in the first and second combustion stages.

A deficiency in the swirl pattern causes a lower combustion temperature, creating what is called a cold spot within the exhaust temperature wheel of the gas-producing turbine.

As the causes of this effect may be related to the above, especially in the aging of the parts, it is advisable to make a prediction of such temperatures.

For this case we will take a different data set from the previous one, obtained from the Mendeley data repository from the following link:

https://data.mendeley.com/datasets/6w3vy3ybhg/3

Esan, Ayodele Benjamin; Ehiaguina, Vincent (2019), “Data-set for Independent Gas turbine for Electricity Generation.”, Mendeley Data, V3, doi: 10.17632/6w3vy3ybhg.3

We’ll predict each exhaust temperature wheel of the gas-producing turbine, using Fuel Gas, Power, Ambient Temperature, Axial Compressor Pressure and Turbine air inlet filter DP as features:

As can be seen, the temperatures corresponding to cans 3 and 4 show a different pattern than the rest.

Conclusions:

From the CO and NOx measurements, behavior patterns can be obtained that indicate some abnormality in the operation of the combustion chambers.

For a better interpretation of the results it is advisable to jointly analyze behavior patterns of related variables such as axial compressor discharge pressure and exhaust temperatures wheel of the gas-producing turbine.

Machine learning algorithms are useful to identify differences in the behavior patterns of variables, as long as the training periods are appropriate.