AAAR 37th Annual Conference October 14 - October 18, 2019 Oregon Convention Center Portland, Oregon, USA
Abstract View
PM2.5 Concentration Prediction Using Convolutional Long Short-Term Neural Network
KAZUSHI INOUE, Ayumi Iwata, Tomoaki Okuda, Keio University
Abstract Number: 165 Working Group: Instrumentation and Methods
Abstract Environmental standard for PM2.5 concentration is provided as a standard value for achieving appropriate protection of human health. Apart from this, there is a provisional guideline as an alerting when it is predicted that the concentration of PM2.5 will be too high. However, existing prediction methods for PM2.5 concentration are often inaccurate over a long period, such as 12 hours ahead, and have not been able to play an effective role in decision making for alerting. Therefore, we tried to predict PM2.5 concentration after 12 hours using deep neural network. We proposed the method of combining the convolutional neural network (CNN) and the long short-term memory (LSTM) to extract the spatiotemporal relationship of the PM2.5 concentration. CNN is a method widely used mainly in the field of image processing, and it is possible to extract image features. In this study, we mapped the concentration data at each air quality monitoring station, reflecting its positional relationship, to generate "PM2.5 concentration image". The spatial relationship between adjacent monitoring stations was extracted by using the CNN method on this image. After extracting the spatial relationship in this method, it was combined with the LSTM. LSTM is a method for handling time series data, and it is possible to extract temporal relationships. Spatiotemporal features at arbitrary points were extracted by inputting time-series data of spatial relationships extracted by CNN into LSTM. These proposed spatiotemporal features were combined with auxiliary data such as date and time, and meteorological data such as wind speed and temperature to generate the proposed model and to predict PM2.5 concentration. In addition, the datasets used to verify the effectiveness of the proposed model were all obtained from Atmospheric Environmental Regional Observation System: AEROS.