Multi-dimentional Time Series Prediction Based On LSTM Neural Netwrok

Since data mining and cloud storage are becoming a hot area of research,there is an increase in the demand for instruction and knowledge from data mining.Time series prediction has a long history.Therefore,how to apply data mining and machine learning to this particular research area has been given a lot of attention.Because multivariate time series have more information than univariate time series.So multivariate time seires predition would have wider application.As neural network is robust and has strong adaptive ability,it has more advantages in multivariate time series prediction.In order to have high precision and computational effeicency,selecting most relevant variable is very important.Secondly,how to select data in the most revelvant time point is as important as selecting most relevant variable.Thirdly,how to select most effective prediction from high-dimension models is very important as well.A prediction algorithm based on LSTM(Long Short Term Memory)neural network has been put forward in this paper and tested with emperical data.In this paper,the main research work are:In order to reduce the dimension of data,this paper presents an dimemsion-reduction algorithm based on Pearson correlation.this algorithm select the most relevant variable according to the Pearson correlations.Because all variables have many time points.The most relevant time point should be determined before model training.This presents an method based on mutual information to determine the most relevant time points.Finally,this paper uses an air quality data which has 9358 samples and 12 variable to compare the effeciency of LSTM neural network with other high-dimension prediction models.This paper trains a LSTM neural network with the variable and time point which are selected from the presented algorithm.According to four different evaluation measures,this paper discusses the effeiciency of support vector machine,Adaboost,BP neural network and LSTM neural network.