PM_2.5 Deep Learning Prediction Modeling Based On Spatiotemporal Attention

Environment protection department need to grasp the concentration of PM_2.5 in a future moment when monitoring.However,the existing PM_2.5 prediction studies only forecast short-term time points,and cannot accurately give the trend of the next period.Existing methods use a lot of prior knowledge,such as distance to represent the PM_2.5 relationship between two places.The diffusion law of PM_2.5 is affected by terrain,meteorological conditions,human activities and other factors.And the actual situation of each site is different,so prior knowledge cannot exhaust all possibilities.In addition,the traditional PM_2.5 prediction method is usually based on time series,which ignores the spatial and temporal relationship between observatories,resulting in insufficient features contained in the data.Based on the model of data,deep learning mines the known or unknown features of prior knowledge,and a series of methods it contains can automatically extract features,which is suitable for solving the problem of PM_2.5 prediction with numerous influencing factors.The main contributions of this paper can be summarized as the following three aspects:1.For the extraction of time features,the prediction method based on circular neural network is studied.The comparison shows that RMSE of the two is close to each other,and LSTM has fewer parameters than Bi-GRU,and the calculation speed is improved by 31%.Therefore,LSTM is selected as the time feature extraction layer.2.For the extraction of spatial features,the feature extractor based on spatial-temporal Attention and ConvLSTM layer is designed,and the prediction model Att-ConvLSTM was built.In order to enhance the robustness of abnormal data during training,insensitive loss function is used instead of the average absolute error function commonly used in prediction model.3.Deconvolution and temporal Attention were used instead of spatial Attention and ConvLSTM.In order to solve the problems of more model parameters and longer training time,the Att-Deconv model with fewer training parameters was proposed.The results show that Att-Deconv training time was only 5%of former training time and RMSE was not decreased,indicating that the temporal and spatial data contain more information.The model accuracy was further improved after adding two Attention mechanisms,indicating that the weight obtained by Attention training conforms to the actual characteristics of data.Compared with the original data,the spatiotemporal data processed by deconvolution reactivated the elements that contributed more to the prediction.4.To test the performance of the model in the real environment,the model presented in this paper was deployed and an application program was written on the raspberry PI embedded platform.The actual data collected by web crawlers were used to predict the change of PM_2.5 within 24 hours,and the results showed that the prediction of the morning low concentration segment by Att-ConvLSTM and Att-Deconv was more accurate than that by LSTM,which proved the robustness of the model in the practical application environment.