Research And Implementation Of PM_2.5 Prediction Based On Neural Network

With the continuous development of smart city and the wide application of IOT technology,the use of big data and machine learning technology to solve practical problems has received widespread attention.This paper focuses on the prediction of PM_2.5 in the construction of smart cities.In the existing related studies,the correlation between pollutants and meteorological conditions has not been comprehensively considered,and only periodic predictions have been carried out for single pollutants.The generalization is poor and there is insufficient accuracy.On the other hand,the prediction of pollutants is not enough granularity,and it is also not enough for refinement,and it is not effective use of existing monitoring data which is monitoring intervals in hours.This dissertation proposes a predictive research method based on different time steps,using previous historical pollutant data and meteorological data to train models with different time windows.Furthermore,using the model to predict the PM_2.5 concentration in the next 1 to72 hours based on the current data of 1 to 72 hours.In order to meet the needs of more massive and complicated air quality data in the future operation of smart cities,a set of Hadoop-based DFS?DFS,Distributed File System?distributed storage systems was designed.The data used in this paper is derived from pollutant data and meteorological data from xing'qing community monitoring stations in Xi'an City in the past five years.Based on various pollutants and meteorological historical data,this paper mainly discusses three prediction models.In order to train the prediction model of different time windows,comprehensively consider the pollutant concentration and meteorological data of the previous time period.Among them,when training predictive models that are larger than one-hour steps,we need to add the prediction results of the one-hour time step model,and so on.The linear regression model describes the linear relationship between pollutants which has a certain degree of accuracy.The BP?Back Propagation?neural network model can fit any nonlinear relationship and can tap the non-linear relationship between pollutants.The input eigenvalues of the model are similar to the linear regression model.Iterative training generates different time window prediction models.The LSTM?Long-short-Term Memory?prediction model uses the memory of the network to describe the delay and continuity of pollutant changes.The input is the previous 1 to 72 hours of data,and output is the PM_2.5 concentration predicted in the next 1 to 72 hours.Based on the historical data of Xi'an Xingqing Community Monitoring Station,this paper compares the effects of the three models,and it is concluded that the use of different prediction models for long out-of-sync predictions can minimize prediction errors.In the stage of model tuning,the optimization schemes of different models were discussed separately.The ridge regression model can further reduce the error of the general linear regression model.Compared with other methods,SGD?Stochastic Gradient descent?optimization scheme can reduce the errors of BP neural network and LSTM network model.Compared with the experimental results,the prediction model of LSTM is less than 1 to 15hours,and the BP neural network model has less error in the 16 to 72-hour prediction step.Finally,a distributed storage system is designed,which is based on Hadoop.Designing multiple nodes is to ensure data scalability and availability.Then building Sqoop tools and Hive tools,whose aim is set up a data warehouse,which can provide supports for further data analysis work.Setting up a Web system to provide predictive services through HTTP services.At the same time,the main five modules of this research system were tested and all achieved the expected results.