We are planning to publish entire or part of projects as educational or showcases for those interested in our services or the way we work.

We start with a power plant NOx CO optimization project. In this scenario the technical analysis has been on a fundamental level. Process simulation has not been considered feasible. Multiphysic simulation of the design is however possible – from gas flow and heat transfer to more advanced simulations of combustion reactions and emission outputs.

NOx emission reduction using Designed Experiments – EON Heleneholm gas fired Power plant – Boiler P12

An 87 run factorial experiment was performed at the EON Helenholm gas fired boiler P12 in order to reduce NOx emissions given a threshold of 100 ppm carbon dioxide (CO).

For the 87 runs, the L81 40 factor three level, 81 row orthogonal standard design matrix was chosen. 6 randomly selected rows were then added for error estimates. Thus, the usual small (8 to 20 run) experiment sequential strategy was rejected.  This due to the customer requirement of a one shot experiment. Natural gas lances where allocated to each factor. During the experiment the lances where rotated. Each of these lances where adjusted to the levels minus sixty, zero, or sixty degrees according to the individual run settings specified in the 87 run designed experiment. The zero setting was the present optimal setting at the start of the experiment.

After running the experiment, the following analysis procedure was performed, to find an optimal operating point:

  1. The experiment data set, was analyzed as a balanced forty dimensional sample, with current operational point as center point. The design generated emission data from 81 gas lance settings (6 replicated), each representing a potential new optimal setting. Thus, due to the L81 matrix orthogonal properties, there will with reasonable certainty, be one or more operational settings with lower emissions. Prior to the experiment the NOx level was 63 mg/MJ. After 87 runs, the level for the best run setting was 48 mg /MJ given allowed CO limit. Thus, a NOx emission reduction of 63 – 48 = 15 mg /MJ was achieved. This by simply adjusting the lance angle settings to the L81 matrix row levels.
  2. Least squares is the most frequently used method when analyzing experimental data sets, including data from L81 designs. Location and dispersion second degree polynomial models where built, one set for the NOx emission response, and one set for the CO emission response. It was obvious that the CO emissions generated dispersion effects indicating stability problems in the combustion process, while the NOx model showed location, or mean value, dependence on a larger set of lances. The instability indicated that, we had a robust design type of problem. Validating the model became a two-step process where: 1) you try to find the lance settings with a low and stable CO level and, 2) decrease the NOx mean value to an optimum low level, preserving stable CO levels within the 100 ppm limit. An approximately disjoint distribution of CO and NOx lance sets, improved the odds for success. It should be mentioned that, finding a statistically valid optimal NOx level point, given a one shot L81, is not possible as the L81 design is heavily fractionated. To “solve” this, we postulated a set of most probable models. Based on these, we gave the operators instructions where to look (in factor space) for an optimum, and based on experience trying to find it. This resulted in a final optimal point parameter setting level of 39 mg/MJ.

The achieved reduction, 63 – 39 = 24 mg/MJ, could, based on the NOx fee system, be calculated to approximately one million dollars a year (2016).

For those of you who are interested in doing your own optimization work, we have attached the data with matrix factor level settings and run measurements.  Each run included the standardized procedure – setting factors, process setting time and measuring time. All runs where made at normal plant operation. The matrix has forty factor columns. All columns are three level columns. The 40 lances are assigned to factor column 1 to 40. Each of the 87 matrix rows specifies the setting of forty factors before each run. And the response vectors to the right are run system response measurements after the run Is completed.

You may note that measurement value 29 and 41 are missing due to combustion process disturbances coinciding with the experiment. However, two missing data in a balanced data set of 87 is of minor importance.

New analysis with JMP 12 shows that a further reduction in NOx emission to a 20 mg/MJ given a CO of 20 ppm could be possible. A power point presentation of initial and updated results are available at this link and calculations of best models at this link. The profiler at the bottom shows 17 mg/MJ and the CO in log10 scale (10 ^1,33 or 20 ppm) with confidence limits.  A “traditional” validation method based on model F ratios, F lack of fit, parameter statistics in combination with a thorough analysis of model residual where used zero residual information content close to zero. The latter use every available data, including process notes, to achieve a low information content. Note that, even lower emission values are possible if you expand the search outside the hyperspace of the experiment. This if you are willing to invest in a new set of test runs.

This, and other results from power plant experiments, open interesting policy perspective for plant owners. It is reasonable to do an optimization of the present hardware, before you decide on investment in new hardware; in this case low NOx burners and/or various flue treatment systems. Otherwise, you do not use the capabilities you already have invested in, you have to invest in larger capacity retrofit processes and increase the risk for secondary problems from the new systems, especially if you add substances such as ammonia, that generate leakage risks not present in the original design. And, in addition, the authorities have a tendency to increase the number of emission to control under increasingly prohibitive requirements. This includes low level process efficiency requirements and high level CO2 level emission limits, both penalized by adding new process equipment. Thus, the wise plant owner would implement, a retrofit policy that the operating process is fully optimized prior to any hardware investment.

The L81 business case shows a designed experiment that, by finding new best operating parameter settings, decreased the plant emission level to a third of the original emission level.