應用類神經網路提升焚化廠風險評估前置資料品質

Abstract

在採用風險評估衡量都市垃圾焚化廠所排放的戴奧辛對周遭生物所造成的影響時，需要各方面正確的資料，特別是焚化廠方所提供的戴奧辛排放數據。另外焚化廠周界大氣戴奧辛濃度檢測的結果往往與風險評估時所使用的ISCST大氣擴散模式模擬結果有明顯差異，造成周界居民與焚化廠雙方的爭執。 本研究將嘗試解決以上兩問題：第一為檢驗焚化廠方所提供的戴奧辛排放數據是否合理；第二為改善因ISCST模式值與實測值間差異造成的問題，兩者皆依資料探勘觀念進行研究。前者採用SOM類神經網路建立焚化廠煙道戴奧辛濃度異常值檢測方法，對異常數據提出合理懷疑。後者採用BPN類神經網路模擬ISCST模式值與實測值之比值，找尋由ISCST模式值推算周界濃度實測值方法。 研究結果顯示，SOM類神經網路在33筆戴奧辛檢測報告中發現4筆異常數據，其焚化廠操作條件與戴奧辛濃度在分群上產生不合理現象，在風險評估上應避免使用。 BPN類神經網路在全國九座都市垃圾焚化廠107筆周界大氣戴奧辛檢測數據中使用90%的數據量做為訓練，10%數據量做為測試，模擬ISCST模式濃度與實測濃度比值，其MSE分別為0.0173及0.0150，效果不佳。若採用相同地域條件的高雄三座焚化廠做BPN類神經網路建置，則可得到更佳的學習效果。最後使用SOM類神經網路進行周界大氣戴奧辛與焚化廠煙道戴奧辛指紋辨識，發現就目前收集到的資料而言，焚化廠煙道戴奧辛與周界大氣戴奧辛大不相同，在此情況下由焚化廠經ISCST模式推算周界戴奧辛濃度的效果不佳。

When using the risk assessment method to examine the impact of dioxins released from municipal solid waste incinerators (MSWIs), we need correct data in all respects, especially the dioxins emission data provided by the operators of incinerators. Furthermore, the dioxin concentrations measured in MSWIs surroundings are usually quite different from those from the result of air dispersion model, such as ISCST. The difference makes it hard for decision makers to issue risk management strategies. In order to address these issues, this study proposes methods of assessing the correctness of emission data and relating ambient concentration measurements to predictions from modeling. The methodologies of this research are developed based on the data mining theory. We adopt SOM to establish outlier analysis method of incinerator flue dioxins concentrations and suggest reasonable explanation to the unusual data. BPN is then used to simulate the ratio of the ISCST modeling value and the measured value, attempting to estimate observed ambient concentrations from the ISCST modeling results. The result of study shows that there are 4 outlier data among the 33 incinerator flue dioxin measurement reports in SOM topology; we should avoid use of the 4 data in risk assessment. In BPN neural network, there are 107 ambient air dioxin measurement reports from the 9 incinerators in Taiwan, and we use 90% data for training and 10% data for testing to simulate the ratio of ISCST-predicted values and the observed values. The MSE values are 0.0173 and 0.0150, respectively, meaning that the relation is not significant. Then we adopt data from 3 incinerators in the same area in Kaohsiung to build BPN neural network and get better result. Finally, we use SOM neural network to identify ambient air dioxins fingerprints and incinerator flue dioxins fingerprints. For the data collected at present, we find that the dioxins fingerprints in the ambient air are quite different from the dioxins fingerprints in the incinerator flues. In this situation, it is not appropriate to estimate the observed value via the ISCST modeling value and BPN neural network.