Please use this identifier to cite or link to this item:
Machine learning to explore the temporal and spatial relationship of PM2.5 concentration changes
Self-Organizing Neural Networks (SOM),PM2.5,Gamma test method,Back Propagation Neural Network (BPNN),
|Publication Year :||2019|
本研究以台灣北部地區作為案例，探討PM2.5的空間和時間特徵有助於判別其可能的污染源。自組特徵映射類神經網路（SOM）可以對高維數據集進行分類，形成有意義的拓撲結構，具有信息提取及可視化的優點。考量到聚類特性和視覺化效果，SOM能夠有效探索高維度多變量空氣污染的相互關係。接著透過分析不同類型站點高污染事件造成的污染原因及來源，比較時空關係的變化，找出與PM2.5濃度變化相關的因素。最後藉由機器學習方法建立模型預測PM2.5濃度。主要結果顯示：（1）人口密度高及交通運輸流量大的地區會造成PM2.5濃度的變化大; （2）PM2.5的濃度變化明顯受到季節遷移的影響; （3）發生長時間大範圍的空氣污染的原因，主要係由於境外霾害所影響;（4）Gamma Test非線性方法的結果是解析PM2.5生成機制的一種方法;（5）不同類型測站受高汙染事件影響程度不同，造成預測PM2.5濃度的準確度有相當大的差異。
In recent years, fine particulate matter (PM2.5) has been a critical air pollutant in many developed countries. Exposure to high concentrations of PM2.5 can cause serious health problems because PM2.5 contains microscopic solid or liquid droplets that are sufficiently small to ingest deep into human lungs. Thus, a precise daily prediction of PM2.5 concentration is notably important to regulatory plans that inform the public and social activities when injury events are foreseen. PM2.5 was legislated as an air pollutant in Taiwan since 2012. However, the generation of PM2.5 is very complicated, including natural and artificial sources. The mix of local sources and regional transportations makes the control and accurate prediction of PM2.5 a very challenge work. Analyzing the spatial and temporal characteristics of PM2.5 could help identifying its possible emission sources. The self-organizing map (SOM) can classify high-dimensional datasets to form a meaningful topological map and has advantages of information extraction and visualization. Bearing in mind the clustering capability and visual interpretation, the SOM is used to explore the interrelationships of high-dimensional multivariate air pollution systems. Analyze the causes and sources of pollution caused by high-level pollution events at different types of stations, compare the changes in time-space relationship, find out the factors related to PM2.5 concentration changes. Finally, establish a model by machine learning method to predict PM2.5 concentration, where the air quality datasets under different time and space in the northern region of Taiwan are used as study case. The main results shows: (1) high density of population and traffic area usually bring high change of PM2.5 concentration. (2) The change of season bring obviously effect. (3) The cause of long-term and large-scale air pollution is mostly due to the damage caused by overseas disasters. (4) The Gamma test method is an effective way to analyze the PM2.5 generation mechanism. (5) The PM2.5 concentration at various monitoring stations are affected by different levels of extreme air pollution events that significantly influences the accuracy of predicting PM2.5 concentration obtained by the machine learning method.
|Appears in Collections:||生物環境系統工程學系|
Files in This Item:
|5.92 MB||Adobe PDF|
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.