Research And Implementation Of PM2.5 Prediction Based On CNNs-GRU Deep Learning

The world has been plagued by air quality issues for a long time,especially for PM2.5 pollution.The traditional statistical models have problems such as low accuracy and few prediction timetseps when predicting PM2.5.Based on the aforementioned issues,this paper proposes a data-driven deep learning model CNNs-GRU.Meanwhile,Wuxi PM2.5 prediction software is designed to finely predict PM2.5 concentration for the next 6 hours.The main work of the paper is as follows:1.This paper analyzes the current research status of PM2.5 prediction with statistical models at home and abroad,and summarizes the existing problems.So a data-driven PM2.5 prediction model CNNs-GRU is proposed,which consists of three parts: CNNs structure,GRU network and fully connected network.Among them,the CNNs structure can automatically extract and fuse the local variation trend and spatial correlation features of multi-station and multimodal air quality data.The GRU network can capture long-term dependency features of time series.Combining the two parts above,they can automatically extract the spatiotemporal features of air quality data and improve the accuracy of PM2.5 prediction.Finally,the fully connected network outputs the final PM2.5 prediction results.2.The Wuxi PM2.5 prediction software based on CNNs-GRU model is designed.Its functions include database design,air quality data acquisition,data preprocessing,prediction model implementation and visual interface design.The database design needs to design the table structure according to the actual requirements of the software,used to store the raw air quality data,preprocessing results,prediction results and other process data,which is the data interaction center of the whole prediction software.The data preprocessing mainly performs numerical coding,missing value filling and so on of the original air quality data,so as to obtain the data reprentation that the prediction model can directly use.The CNNs-GRU model implementation is the core of the whole prediction software.This paper uses the deep learning framework Keras to model,and adopts some optimization methods of BatchNormalization,Earlystop,etc.Finally,the visual library Echarts and Baidu Map API are used to design a web interface to display PM2.5 prediction results in real time.3.This paper builds test environment and conducts functional tests on each functional unit of air quality data acquisition,data preprocessing,prediction model implementation and visual interface design of Wuxi PM2.5 prediction software.And we compare and analyze the prediction performance of the proposed CNNs-GRU model and other six models of ARIMA,ANN,LSTM,APNet(LSTM/GRU)and RNN-CNN.On the two test sets of Wuxi in January and June 2017,the prediction accuracy(100-MAPE)of CNNs-GRU model was as high as 76.902% and 70.053% respectively,which was at least 6.6% and 5.1% higher than the other prediction models.The RMSE index is 15.570 and 9.577,the MAE index is 10.720 and 7.068,both of which are significantly better than the other six prediction models.