Predictive Emission Monitoring (PEM) Systems Development and Implementation

Transcription

1 19 TH SYMPOSIUM OF THE INDUSTRIAL APPLICATION OF GAS TURBINES COMMITTEE BANFF, ALBERTA, CANADA OCTOBER 17-19, IAGT-102 Predictive Emission Monitoring (PEM) Systems Development and Implementation K.K. Botros NOVA Research & Technology Center and C. Williams-Gossen, S. Makwana, L. Siarkowski TransCanada Pipelines Ltd. October 17, 2011

2 Outline 1. Motivation. 2. Fundamentals of NN-based PEM Models. 3. Application to RB211 Engines. 4. Application to LM2500 Engines. 5. Application to LM1600 Engines. 6. Comparison with AP-42 Emission Factors. 7. Discussion. 2

3 Outline Motivation 1. Since 1968 the Environmental Protection Agency (EPA) has published emission factors for gas turbines in its Compilation of Air Pollutant Emission Factors called AP-42*. 2. Government regulatory agencies, industry and others use this document to estimate emissions of atmospheric pollutants, a critical step in the development of effective air quality management strategies. 3. In Canada, actual NOx emissions from the pipeline industry are generally not accurately determined since CEM (Continuous Emission Monitoring) systems are not required on compressor drivers. 4. For units post 1992 CCME** guidelines, there is a measurement and monitoring requirement, but CEM is not required for pipeline units. * EMISSION FACTOR DOCUMENTATION FOR AP-42 SECTION 3.1 STATIONARY GAS TURBINES Prepared for: Office of Air Quality Planning and Standards-U.S. Environmental Protection Agency Research Triangle Park, NC, Prepared by: Alpha- Gamma Technologies, Inc Falls of Neuse Road Raleigh, North Carolina, April ** National Emission Guidelines for Stationary Combustion Turbines; Canadian Council of Ministers of the Environment, 1992; ISBN:

4 Outline Motivation 5. As an alternate method to quantify the NOx emission from smallto-medium size gas turbines typically used in pipeline systems, a PEM (Predictive Emission Monitoring) system was developed to provide a better and a credible estimate of NOx emissions at all loads at prevailing ambient conditions. 6. Once integrated on a yearly basis, a PEM system provides a better estimate of the NOx emissions inventory compared to estimation methodologies that rely on the AP-42 emissions factors. 4

5 Outline Motivation 6. TC has pioneered work on PEM s development over the past 5 years. 7. PEM systems were developed for RB211, LM2500 and LM1600 engines. 8. PEM s were implemented and evaluated at 3 pilot stations (Winchell Lake, Clearwater and Hussar). 9. This presentation provides the fundamental basis for these PEM systems, implementation techniques and results. 5

6 Outline NN-based PEM Model Step 1: CEM measurements during four different seasons of a year. Step 2: Design and optimize NN architectures, trained on these CEM data. Step 3: Arrive at optimum architecture Multi-Layer Perceptron NN with one hidden layer of two neurons. Outcome: PEM model written in C++ or FORTAN Implemented in Compressor Equipment Health Monitoring (CEHM). Produces NOx data at any time intervals. 6

7 Winchell Lake Winchell Lake C/S Measurements Clearwater 1. CEM: NO x, NO, CO, CO 2, O 2 and moisture in the exhaust stack are measured by a portable CEM system using instruments and on-site calibration techniques approved by the respective U.S. EPA, i.e.: Carbon Monoxide: EPA Method 10 Oxides of Nitrogen: EPA Method 7E (Typically <2% of Calibration Span) Carbon Dioxide: EPA Method 3A Oxygen: EPA Method 3A Total Hydrocarbons: EPA Method 25 Moisture content: Alberta Environment Method 4 7

8 Hussar Winchell Lake C/S Measurements 2. Pertinent engine performance parameters and ambient conditions at the site are obtained from the station/unit data. 3. These parameters are: Pamb, Tamb, (N1,N2,N3), P3, Q f, W s. 4. Gas samples are also taken to determine the composition of the fuel gas to determine the HHV. 8

9 Winchell Lake C/S Measurements Stack Testing System (CEM) Courtesy of Air-Tec Consulting Ltd. 9

10 10

11 11

12 12

13 13

14 14

15 Multi-Layer Perceptron Architectures Activation function: a) logistic sigmoid b) hyperbolic tangent 15

16 Multi-Layer Perceptron Architectures 16

17 Multi-Layer Perceptron Architectures Profile : MLP 6:6-2-1:1, Index = 2 Train Perf. = , Select Perf. = , Test Perf. = Selected Architecture: Tamb N1 N3 P3 NOx Qf Ws Hidden Layer Ranking Tamb N1 N3 P3 Qf WS Correlation Coefficient Error S.D. (ppm)

18 Field Validation (Winchell Lake) In 2008, the PEM model was installed on a C/S CEHM (SCADA) system. Collecting minutely (every minute) data since June CEM measurements were conducted on Sept 12, 2008 at varying engine loads (9,956 kw 18,087 kw). CEM was compared to PEM. 18

19 Nox (ppm dry) Measured vs. PEM (RB211-24C) Cem Measurements PEM Prediction Error Power (kw) 19

20 Winchell Lake C/S Measurements LM2500 Development Clearwater Unit #1 20

21 Winchell Lake C/S Measurements CEM Measurements (Clearwater) Conducted two CEM tests on two different days in 2009: 1. March 24, 2009: 191 data points Ambient Temperature + 1 o C Power varied from MW PT Speed varied from RPM 2. August 18, data points Ambient Temperature + 24 o C Power varied from MW PT Speed varied from 3880 to 5260 RPM 21

22 Winchell Lake C/S Measurements CEM Measurements (Clearwater) From CEM: NOx in ppm (dry-15% O2corrected ). O2% (dry). From CEHM: Fuel Gas Flow Rate. Fuel gas mixture composition. We can then Calculate: NOx emission in kg/hr. EF = kg/hr / (Qf * HHV) = kg/gj 22

23 Multi-Layer Perceptron Architecture NOx (ppm) Ranking Correlation Error S.D. Tamb N1 P3 Qf Coefficient (ppm) NOx (kg/hr) Ranking Correlation Error S.D. Tamb N1 P3 Qf Coefficient (kg/hr)

24 NOx (kg/hr as NO 2 ) NOx (ppmv-dry 15% O2) Measured vs. PEM (LM2500) 250 NOx (measured) 200 NOx (Predicted by PEM) NOx (measured) NOx (Predicted by PEM) Fuel Flow Rate (m 3 /hr) Fuel Flow Rate (m 3 /hr) 24

25 Winchell Lake C/S Measurements LM1600 Development Hussar 25

26 Winchell Lake C/S Measurements CEM Measurements (Hussar) Conducted two CEM tests on two different days in 2009: 1. June 25, 2010: 188 data points Ambient Temperature +26 o C Power varied from MW PT Speed varied from RPM 2. October 26, data points Ambient Temperature -2 o C Power varied from MW PT Speed varied from 5634 to 7014 RPM 26

27 Multi-Layer Perceptron Architecture NOx (ppm) Ranking Correlation Error S.D. Tamb N1 P3 Qf Coefficient (ppm) NOx (kg/hr) Ranking Correlation Error S.D. Tamb N1 P3 Qf Coefficient (kg/hr)

28 NOx (kg as NOx/hr) NOx (ppmv- dry-15% O2) Measured vs. PEM (LM1600) Measured NN Predicted by PEM Measured NN Predicted by PEM Fuel Flow Rate (m 3 /hr) Fuel Flow Rate (m 3 /hr) 28

29 Uncertainty Analysis Uncertainty in Level of Uncertainty (+ or -) Error in NOx (ppmv + or - %) Error in NOx (kg/hr %) Ambient Air Temperature 1 deg C 0.90% 3.50% GG Speed (N1) 1% 3.00% 6.00% Compressore Discharge Pressure (P3) 1% 2.50% 5.30% Fuel Consumption (Qf) 1% 2.80% 5.90% 29

30 EPA, AP-42 and CCME AP-42 emission factors of NOx for an uncontrolled stationary GT are: All Loads: EF = kg/gj (0.295 lb/mmbtu) based on HHV Loads greater than or equal to 80%: EF = kg/gj (0.32 lb/mmbtu) based on HHV CCME 1992 Guideline, ISO rating, CONTROLLED NOx level requirements are: 500 g/gj output for power < 3 MW 240 g/gj output for 3 MW < power < 20 MW 140 g/gj output for power > 20 MW 30

31 NOx (g of NOx/GJ of Output Work) NOx (kg as NOx/GJ) Comparison with AP-42 and CCME EF = kg/gj (LM1600) 0.10 EF = kg/gj Fuel Flow Rate (m 3 /hr) CCME Emission Intensity = 240 g/gj (for ourput power range of 3-20 MW) Shaft Power (kw) 31

32 NOx (g/gj output work) NOx (kg as NOx/GJ) Comparison NN-Based Between PEM Model RB211, LM2500 and LM From Stack PPM and O2% Measurements RB211 LM2500 LM EF = kg/gj EF = kg/gj 700 RB LM Fuel Flow Rate (m 3 /hr) CCME Emission Intensity = 140 g/gj (for ourput power > 20 MW) Fuel Flow Rate (m 3 /hr) 32

33 80% Load NOx (kg As NOx/GJ) NOx (kg as NOx/GJ) 80% Load One Year Data Winchell Lake Clearwater (unit 1) EF = kg/gj 0.15 EF = kg/gj 0.1 EF = kg/gj 0.1 EF = kg/gj Fuel Flow rate (m 3 /hr) Fuel Flow rate (m 3 /hr) Winchell Lake: RB211-24C Clearwater: GE LM

34 NOx (kg as NOx/GJ) 80% Load One Year Data Unit 6 One Year Data Unit 7 One Year Data Series5 EF = kg/gj EF = kg/gj Fuel Flow Rate (m 3 /hr) Hussar: Units 6 and 7 (LM1600) 34

35 Comparison (One Year Data) Winchell Lake RB211-24C Total Running Time in 1 Year (hours) 5950 PEM Predicted NOx Emission (tonnes of NOx) NOx Emission Based on AP42 Factors (tonnes of NOx) Clearwater LM2500 Total Running Time in 1 Year (hours) 6063 PEM Predicted NOx Emission (tonnes of NOx) NOx Emission Based on AP42 Factors (tonnes of NOx) Hussar LM1600 Unit 6 Unit 7 Total Running Time in 1 Year (hours) PEM Predicted NOx Emission (tonnes of NOx) NOx Emission Based on AP42 Factors (tonnes of NOx)

36 Conclusions 1. NN-based PEM model is simple and does not need to include multiple hidden layers and multiple neurons. 2. Correlation parameter > 0.99 was achieved with MLP architectures. 3. Error St. Deviation of Trained NN based PEM model is ~ 4.85 (ppmv) and 1.32 (kg/hr). 4. NOx emission inventory determined by the developed PEM models provide better estimates than AP-42 emission factors for the three tested engines (RB211-24C, LM2500 and LM1600). 5. Close correlations between field test results and the PEMS predictions are demonstrated. This demonstrates that PEMs are viable alternative to CEMs, and produce more realistic results than using AP-42 factors. 36

37 Acknowledgement TransCanada: Caroline Selinger, Mark Cheung, Jim Cormack Mike Lee, John Met, Juan Velasquez (Implementation) Lyle Hartman (Winchell), Gary Blick (Clearwater), Denis Burrows (Hussar) Nicholas Kertesz, Jordy Miller, Marilyn Carpenter, Mark Blundell Anthony Tse, Thomas Robinson Maxxam Analytic Inc. Sean Miner, Dennis Skwarchuk, Reg Mullett, Daryl Ford, Darcie Thauvette 37