Research On Complex Time-series Anlysing Based On Deep Learning

Time-series analysing means feature extration and modeling for a certain period of time-sereis data.It has been widly used in many fields such as macro-economics,astronomy,oceanography and medical science.Following the age of big data,the time-series data nowadays are more large in both number of dimensions and length compared with former time-series data.Which on one hand makes it possible to use more complex models to handle the problems,but on the other hand requires the algorithms be more efficient.Besides,the traditional methods for time-series analysing might not work well result from the growing complexity of time-series data,more complex models like deep neural networks may be suitable for this case.In addtion to the growing of complexity and length,masssive data often comes with missing observations,how to repairing time-series with missiong observations or directly learning from time-series with missing observation become meaningful questions.From simple to complex,this paper firstly proposes an anytime methods for time-series prediction with missing observations.Theoretical proof and experimental results will be described to show the efficiency of the methods.Then,after finding the facts that the traditional methods can’t handle complex time-sereis well from comparison experiments,this paper proposes complex timeseries prediction model,which is based on conditional random field and recurrent neural network.This model can be used as upper layer predicter in stacking process or be trained using deep learning methods.Using the experimental results in real data,the efficiency of the model is verified.Last but not least,this paper focus using the mechanism of generative adversarial networks to deal with the complex time-series repairing problem.First,this paper improve SRGAN,and then proposes a complex time-series repairing model and uses the trianing processes of generative adversarial networks to train the model.Using experimental results from real-world industrial time-series,this paper shows the model’s effciency.