Research On Time Series Prediction Schemes Based On Neural Network Methods

Time series data widely exist in many fields such as finance,astronomy,industry,medicine,electricity and so on.Time series has three characteristics: continuity,randomness and periodicity.These characteristics show the feasibility and difficulty of time series predicting.Time series predicting has a wide range of applications,including financial market forecasting,power demand forecasting,environmental weather forecasting and medical diagnosis.It has great value whether in order to obtain commercial benefits or avoid risks for improving the accuracy of time series predicting.Time series predicting problem can be divided into fine-grained prediction and coarse-grained prediction according to time scales predicted.At present,the main prediction methods include statistical learning methods,traditional machine learning methods,feed-forward neural network methods and recurrent neural network methods.These methods have their own limitations or need to be combined,adjusted or improved according to actual application scenarios.Even little improvement in prediction accuracy can bring huge benefits in actual application fields.Neural network has general approximation ability,strong calculation ability and expression ability.It is an ideal rule learning machine and pattern learning machine and can be used to develop higher-level predictors.This paper is devoted to the study of the working principle of neural network,as well as the adjustment and improvement of its structure,and the combination of neural network models,etc.,so as to establish coarse-grained and fine-grained prediction models to improve the accuracy of time series predicting.To solve the problem of time series fine-grained prediction,this paper proposed a modified model based on LSTM called LSTM-corr.LSTM-corr is a feed-forward neural network correction layer added to LSTM to correct LSTM results,thus providing more accurate and stable prediction results.The data center energy consumption prediction is taken as a specific application scenario,and the data center energy consumption series generated by the data center simulator GreenCloud is taken as an experimental data set.The experimental results show that the prediction effect of LSTM-corr model is significantly better than that of a single LSTM model.To solve the problem of coarse-grained prediction of time series,this paper proposed two schemes: trend prediction scheme based on neural network language model and coarse-grained prediction scheme based on auto encoder.The former is coarse-grained range prediction and the latter is coarse-grained specific value prediction.The accuracy of both proposed predictions schemes is higher than that of traditional algorithms.