許多研究人員使用機器學習技術來提高其空氣污染濃度預測模型的準確性，讓人類可以事先得知準確的空氣汙染資訊，以避免直接曝露於汙染的環境中。據我們所知，目前並沒有研究對大範圍區域的汙染來源進行預測。為準確定位汙染來源，本研究建立了空氣污染來源識別系統，稱為Air Pollution Source Identification System (APSIS)，其使用tensorflow建立三種神經網路的分析模型來找出空氣汙染來源。APSIS在相對較小的區域內收集數據，例如空氣污染濃度，風速和風向。並預先處理收集的數據，目的是確定污染物分佈是正確的，以防止APSIS受到異常值和其他不穩定因素的嚴重影響，如風向。使其能更準確地找出汙染來源。之後，使用高斯擴散模型進行污染物擴散模擬，並與實際擴散情形進行比較，確認高斯模型的準確性是否可以應用在污染來源辨識中，並比較三個神經網絡模型在空汙染來源辨識中的準確度，最後提出一種最適用於識別空氣污染源的模型。 Many researchers use machine learning techniques to enhance accuracies of their air-pollution concentration prediction models so that people can acquire accurate information in advance to avoid exposing themselves in this polluted environment. To the best of our knowledge, currently, there is no research which identifies air pollution source in a wide area. To accurately locate pollution sources, in this research, we create an air-pollution identification system, called Air Pollution Source Identification System (APSIS), which adopts tensorflow to establish three neural-network-based analytical models with which to find pollution sources. The APSIS collects environmental data, such as air pollution concentration, wind speed and wind direction, in a relatively smaller grid area. Next, collected data are tuned when necessary to prevent the APSIS from being seriously affected by outlier and other unstable factors, like wind direction. The purpose is to identify pollution distribution and then more accurately find out the sources. After that, the Gaussian diffusion model is used to simulate the diffusion of pollutants, and compared with the actual diffusion situation, to confirm whether the accuracy of the Gaussian model can be applied to the identification of pollution sources. Then compare the accuracy of three neural network models in the identification of air pollution sources, and finally propose a model that is most suitable for identifying air pollution sources.
