Search

The emissions from coal power plants have serious implication on the environment protection, and there is an increasing effort around the globe to control these emissions by the flue gas cleaning technologies. This research was carried out on the limestone forced oxidation (LSFO) flue gas desulfurization (FGD) system installed at the 2*660 MW supercritical coal-fired power plant. Nine input variables of the FGD system: pH, inlet sulfur dioxide (SO2), inlet temperature, inlet nitrogen oxide (NOx), inlet O2, oxidation air, absorber slurry density, inlet humidity, and inlet dust were used for the development of effective neural network process models for a comprehensive emission analysis constituting outlet SO2, outlet Hg, outlet NOx, and outlet dust emissions from the LSFO FGD system. Monte Carlo experiments were conducted on the artificial neural network process models to investigate the relationships between the input control variables and output variables. Accordingly, optimum operating ranges of all input control variables were recommended. Operating the LSFO FGD system under optimum conditions, nearly 35% and 24% reduction in SO2 emissions are possible at inlet SO2 values of 1500 mg/m3 and 1800 mg/m3, respectively, as compared to general operating conditions. Similarly, nearly 42% and 28% reduction in Hg emissions are possible at inlet SO2 values of 1500 mg/m3 and 1800 mg/m3, respectively, as compared to general operating conditions. The findings are useful for minimizing the emissions from coal power plants and the development of optimum operating strategies for the LSFO FGD system.

Like most nanomanufacturing processes, molecular beam epitaxy (MBE) processes are based on atomic-level control of growing films and thus are sensitive to subtle changes that make repeatability and reproducibility of desired performance indicators a nontrivial task. The gamut of challenges include insufficient understanding of atomic-level interactions, involvement of a large number of candidate process variables, lack of direct observation and measurement techniques for key performance indicators, and significant cost and time requirements for conducting experiments. A conventional design of experiment-based analysis becomes an unrealistic option due to its demand on extensive experimentation. In this paper, we present a hybrid approach that combines current process knowledge, artificial neural networks, and design of experiments (DOE) to make use of preliminary experimental data to analyze the process behavior, enhance process knowledge, and lay down foundations for cost effective systematic experimentation. Based on preliminary experimental data generated while exploring the MBE process for growing a MgO interface layer on 6H-SiC substrate, we developed a neural-network-based meta model that can interpolate and estimate the process responses to any combination of process variable settings within the input space. Using the neural-network model trained on preliminary experimental data, we estimate the process responses for a three-level full-factorial DOE runs. Based on these runs, the DOE based analysis is carried out. The results help explain the MgO film growth dynamics with respect to process variables such as substrate temperature, growth time, magnesium source temperature, and trace oxygen on the initial substrate surface. This approach can be expanded to statistically analyze the dynamics of other complex nanoprocesses when only the exploratory preliminary experimental data are available. This approach can also lay the foundation for efficient and systematic experimentation to further analyze and optimize the processes to address issues such as process repeatability and reliability.