The Gas Prediction of Biodegradable MSW with MSWI Ashes Addition by Using Backpropagation Neural Network

• Main Authors: Chao-Chuan Yang, 楊朝全
• Other Authors: Huang-Mu Lo
• Format: Others
• Language: zh-TW
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/49kt49

碩士 === 朝陽科技大學 === 環境工程與管理系碩士班 === 94 === Municipal solid waste (MSW) treatment has been transferred from landfill to incineration associated with composting and recovery and recycling due to the lesser available land for landfill in Taiwan. However, the residues such as bottom ash and fly ash generated still account for a volume and weight ratio by up to 10 and 25%, respectively. Thus, the treatment and disposal of MSW incinerator (MSWI) ashes become another environmental issue and needs further treatment to prevent secondary pollution. MSWI ashes have been practiced for landfill cover for many countries including Taiwan. However, the reaction mechanisms of co-disposal are not fully clear and needs a theoretical and experimental investigation for a better understanding of baseline information to meet the practice requirement. This study examined the possible utilization of MSWI ashes in anaerobic bioreactors. In particular, using the experimental results such as gas generation rate and metals release from bioreactors to train and predict the trend by backpropagation network (BPN) is the major focus of this study. Results showed that bottom ash added ratio of 100 g l-1 and fly ash added ratio of 10 and 20 g l-1 has the potential to enhance the gas generation rate. This phenomenon brings the advantage of MSW faster biostabilization and potential energy recovery. The input parameters chosen were pH, conductivity, salinity, total solid, volatile solids, chemical oxygen demand, alkalinity, volatile fatty acids, microbes etc. The outputs selected were gas generation rate, Ca, K, Mg and Na. In order to optimize the predicting results, gas accumulation in control bioreactor (blank1) was used to train the learning number and to analyze the values of root mean square (RMS). Results showed that the stability could be obtained after 3500 training times. Thus, the training number was chosen as 5000 for the following modelling. In addition, the addition of related coefficient (R) greater than 1.2 was another screen condition to eliminate the insufficient data from the training and verification bioreactors. These screening conditions thereby resulted in the generation of suitable hidden layers and learning speed for the predicting modelling of BPN. The results of modeling in gas accumulation and alkali metals release were in a good agreement with the experimental results. The R values exceeded 0.95 and showed a high linear relationship. RMSs fell below 100 except the accumulation of Ca ions in the 20 g l-1 bioreactor. In the modelling of gas production per week and Ca release, all average Rs were above 0.8 and all RMSs were below 35 except the medium-high relationship in the 100 g l-1 bottom ash added bioreactor, Ca release in verification set of blank1 bioreactor and in training set of blank 2 bioreactor and verification set in 20 g l-1 fly ash added bioreactor. From these results, it is noted that prediction modeling was found better in gas accumulation than in gas production per week. Particularly, the output values by BPN model were closed to that of the experimental bioreactors. These phenomena indicated that suitable Ca release could enhance the gas generation rate which has been found in the ashes added bioreactors than in the blank ones in the first stage of MSW digestion.