Time-series Imputation And Prediction Usinging Gate Recurrent Units And Generative Adversarial Networks

Recent advances in sensors,cloud computing and Internet of Things,massive time series data have been constantly generated and recorded.The mining and analysis of these time series data has brought great convenience to our daily lives.Such as weather forecasts,climate forecasts,stock price forecasts,commodity retail forecasts,etc.In view of the wide application of time series,many time series prediction methods have been proposed,and most of them use the auto-regressive generative mode,where historical data is provided during training and replaced by the network's output during inference,causing the model to accumulation errors in inference,thereby reducing the prediction performance of the model.In addition,missing data is a frequent occurrence due to sensor failure or improper transmission,storage or another reasons.There is a strong dependence between the historical value and the current value within time series data.Therefore,effective missing value processing methods can promote advanced forecast of time series data.Most of the traditional missing value processing methods do not consider the dependency of time series data,so it is difficult to achieve accurate filling effect.Aiming at the limitations of current autoregressive time series prediction models,this paper combines the idea of GAN and proposes a time series prediction model based on GAN and GRU.The model uses a gated recurrent neural network as a generator to learn the distribution characteristics of time series data,and uses a discriminator to improve the generator's sequence-level prediction performance.After adversarial training,this model can generate a time series that conforms to the original data distribution.Data prediction value.Aiming at the missing values of time series data,after fully studying the characteristics of time series data,this paper proposes a model GAN-BI to learn the distribution characteristics of time series data from historical data to current data and from current data to historical data.The model GAN-BI is based on the idea of de-noising autoencoder and generative adversarial networks.After training,the model can directly fill in the missing value on the original data set.The experimental results under the real data set prove the effectiveness of the time series data prediction method and the time series data missing value filling method proposed in this paper.