The Research Of Time Series Prediction Based On GRU Neural Network

Time series is a series of values observed in the time dimension,and it is a special form of stochastic process.In our real life,there is a large amount of data based on time,such as weather,finance,industry,agriculture,transportation and all walks of life.The time series prediction is an important research in the field of big data analysis and data mining,which analyses and mines through time-series data collected in the past,and finds some rules to predict future data.Time series prediction has great significance in various fields,it can better planning the future development and reducing unnecessary losses.Time series prediction relies on multiple disciplines and it is a crossover study of multi-domain.In the earlier of time series prediction,the method of mathematical statistics is used to predict the time series by using quantitative or qualitative forecasting methods.Generally,the traditional prediction efficiency is poor and the accuracy is low.In recent years,with the development of data mining and machine learning,people begin to realize that artificial neural network has good performance in time series prediction.Using the recurrent neural network could be good for training time series which exists a certain rules,and get the predicted value of time series.Inthis paper,the mainly done for prediction of timing data as follows:1.The paper describes the background and current situation of time series prediction,and analyzes the shortcomings of the traditional mathematical statistical prediction method and artificial neural network.2.Gated Recurrent Unit neural network,which is the optimized structure of recurrent neural network,is used to research of time series prediction.In the recurrent neural network,there are the problems of long-range dependencies and gradient disappearances.So the selection of the GRU neural network,which is less of threshold structure and more efficient.3.Second exponential smoothing is used to correct predicted data after the GRU neural network forecast the time series data.And the prediction result with higher prediction rate was obtained.4.The input training data would be processing of dimension,so it would easy to find more hidden data and train a neural network better.5.Four different sets of data were used for experimental verification,and three different analysis methods were used for error analysis.The experimental results show that the forecasting scheme proposed in this paper prediction accuracy is higher.The research in this thesis improved the accuracy of time series prediction,and it is very important to the industrial production and the time series analysis in real life.