| Summary: | 碩士 === 國立中山大學 === 環境工程研究所 === 101 === Lowering the organic concentration in raw water at water treatment plants can reduce the growth and reproduction of microorganisms in water distribution system, because in the phenomenon of “post-growth” or “re-growth,” the organic carbon is a key element in microorganism growth. Thus, if it is possible to treat and predict the low concentration of organic carbon in treatment water which is necessary for microorganisms to grow, it is possible to effectively control the post-growth of microorganisms in the water distribution pipes.
Neural networks are important in artificial intelligence; it is a kind of applied technology that has become a subject of popular research and development. Neural networks are like human brains, which exhibit the abilities to learn, recall, summarize, and deduce based on samples or data training.
This study focuses on the changes in concentration of assimilable organic carbon (AOC) in two high-grade water treatment plants (represented as A and B), using SPSS 17.0 and neural network system AutoNet (6.03) suite programs to conduct linear regression analysis and correlation analysis to explore water quality changes in AOC concentration with the sequences at water purification plants. Conclusions show that different water sources would result in various AOC concentrations in the treatment water of water treatment plants. Plant A’s treatment process needs to remove at least 50% of the total AOC to reduce the output concentration to 50 µg Acetate-C/L in water, but Plant B only needs to remove 6.7% of the total AOC to reduce the output concentration to 50 µg Acetate-C/L.
This study finds that there are distinctions in the predictive models computed by different statistical software. Results present that the predictive model derived from the stepwise regression analysis was worse than that derived from the enter method and AutoNet in terms of error values and error ratios, while the enter method and AutoNet method showed similar error values and error ratios. However, in terms of R value correlation, the enter method was better than the AutoNet method, but the R values in predictive models established by these three methods were showing all high correlation.
|
|---|
