Application of artificial neural network for air quality forecast and anomaly detection of intelligent air quality sensors

• Main Authors: Huang, Yen-Chi, 黃彥齊
• Other Authors: Bai, Hsun-Ling
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/y36sc4

碩士 === 國立交通大學 === 環境工程系所 === 106 === Artificial neural network(ANN) is a mathematical model that can be used to solve the problem such as classification and regression. This study is divided into two parts: The forecast of PM2.5 concentration of air quality monitoring station and the anomaly detection of air quality sensors network based on ANN. For the air quality forecast, the ANN models were trained with data from November 2013 to April 2014, then the data from November 2015 to April 2016 and November 2016 to April 2017 were used to test the trained models. The target station of forecast is Xitun. The features suitable for forecast was selected first. Then long short term memory network(LSTM) and back propagation neural network(BPN) were used to forecast the PM2.5 concentration of next 1 to 4 hours in Xitun. The results showed that the overall accuracy of LSTM was better than that of BPN. The R2 values of forecast were from 0.92(1hr) to 0.66(4hrs), which decreased with longer forecast interval. The accurate rate of forecast on whether the concentration exceed the air quality standard(>35.4 g/m3 or >54.4 g/m3) was higher than 84%. In terms of anomaly detection, the sensor data from NCTU(National Chiao Tung University) and CTSP(Central Taiwan Science Park) from May 2017 to May 2018 were taken as the research subject. The data would be detected as anomaly if the gap between prediction and measurement of target sensor was higher than the threshold. And for CTSP sensors, some conditions were added to separate the anomaly to different pollution sources. The results showed that the method proposed in this study can effectively detect the failure of sensors and identify the pollution source within 90 minutes of pollution occurrence. Most of the pollution in CTSP in April 2018 came from the sea area.