Long short-term memory - Fully connected (LSTM-FC) neural network for PM2.5concentration prediction
•The LSTM-FC neural network can give an accurate prediction of urban PM2.5 contamination over the next 48 hours.
•The LSTM-FC neural network can handle the long-range dependence of PM2.5 contamination.
•The LSTM-FC use a fully connected neural network to combine the spatial information of surrounding stations.
Abstract：long short-term memory - fully connected (LSTM-FC) neural network
using:historical air quality data, meteorological data, weather forecast data, and the day of the week
effect:
1、model the local variation of PM2.5 contamination
2、a neural network-based spatial combinatory to capture spatial dependencies between the PM2.5 contamination of central station and that of neighbor stations
如何解决空间问题：
note that the LSTM-based temporal simulator only exploits the temporal information. Then the predictions of the central station and sur-rounding stations given by LSTMs are entered into the fully-connected neural network-based spatial combinatory. The final output is the final prediction of the central station.
如何解决时间问题：
The goal of our model is to forecast the PM2.5 over 48 h, spe-cifically, for the next 1~6 h, we predict the real-valued concentra-tion of PM2.5, and with respect to 7~12, 13~24, 25~48 h, we predict the maximum and minimum concentration for each time interval respectively, as seen in Fig. 1(a). To make a forecast, some related data in the past hours is entered into the model, so our model is a multi-step-ahead forecasting model as seen in Fig. 1(b).
创新点：
1) The temporal simulator and spatial combinatory handled temporal information and spatial information, respectively, leading to an effective and interpretable prediction result. The LSTM-based temporal simulator fed with all local historical data which could handle the long-range dependencies. Meanwhile, the spatial combinatory com-bined the predictions of the surrounding temporal simulators.
(2)We combined the spatial information of surrounding stations ac-cording to geographical distance, instead of using data from all stations in Beijing, which averted over-fitting and mitigated thecomputational challenge.
结果
