Research On Forecasting Method Of Air Quality Time Series Data

Environmental air quality is related to people's life closely,air pollution has become a major concern of human beings with the rapid development of economy.Accurate prediction of air quality can help people to arrange their own life effectively and provide a basis for the governance of human living environment.However,with the accumulation of a large number of air quality data,we need new technology to analyze the high dimensional air quality data.In this thesis we put forward a new prediction method of time series data by combining empirical mode decomposition method,One-Hot coding method,principal component analysis method with feedforward neural networks,The new method is used to predict the time series data of two standard data sets of air quality in Beijing,namely Beijing PM2.5 data set and Beijing Olympic Sports Center air quality data set.First of all,the experiments decompose the time series data to be predicted in the original data set with empirical mode decomposition method to obtain several intrinsic mode functions and a residual item,based on them,the subsequent neural network training experiments can be effectively completed.After the model is successfully trained when the minimum error of the neural network is 0.001 by adjusting the weight for many times,we input prediction data set into the training model,and obtain the final prediction result by adding the predicted results linearly.Secondly,In order to make the prediction results more accurate,we use One-Hot coding method to encode the categorical variables in the original data set that cannot be directly used for neural network to obtain binary data set that can be used to train neural network.Then,the experiments reduce the dimensions of the external variables excluding the time variables and the time series data to be predicted in the data set with principal component analysis method.After several times experiments,the number of variables used in this experiment was determined.At the same time,binary data set obtained by One-Hot coding,the data set which is obtained by principal component analysis method are spliced with each intrinsic mode function including trend,to get the same data set as the total number of intrinsic mode function and residual item.After all,we input the standardization of data set into neural network training to get all the training model to predict the prediction set.Finally,ARIMA(Auto-Regressive Integrated Moving Average)model was used as comparison model to simulate and predict the predicted data sets.For the prediction indexes PM2.5 and PM 10,it is found that the mean square error of the new model is much smaller than that of the ARIMA model by calculating the mean square error of the prediction results and the true values of the two models.This shows that better prediction results can be obtained by processing experimental data set with One-Hot coding method,principal component analysis method,empirical mode decomposition method,Data standardization method and adding the prediction results up after neural network training.The prediction model based on the proposed method is superior to the ARIMA model.