Time Series Forecasting Based On Recurrent Neural Networks

Time Series is widely applied in various fields,such as industry,marketing,so-ciology,environment,and education,etc.By means of analyzing and mining time series data,we can build up forecasting models to predict upcoming data.Through time series forecasting,we can provide users with various proactive and instructive suggestions for reacting to the changes in the future.Meanwhile,we also can discover the latent regularity embedded in the time series data so that we can understand related scientific theories and social phenomenons better.However,due to the specific attributes of time series and the new characters of currently emerging time series data,there are some challenges facing us.On the one hand,different from general data,the specific attributes of time series give rise to two problems.First of all,time series are characterized by the time order,which means the data points in it own high relevance,of which we should take full account.In addition,in view of the different lengths of time series and the restriction of Markov property,it is necessary to utilize a sliding window to construct samples for training and testing,but in practice,it is difficult to choose an appropriate window size.On the other hand,different from the traditional time series analysis,there are some new characters of cur-rently emerging time series data,which bring new challenges for us.Firstly,with the development of technology,there are massive time series data generating all the time,thus the efficiency of the forecasting model should be taken into consideration.Sec-ondly,the sampling rate of these time series tends to become higher and higher(such as hours,minutes,even seconds),which introduces a large number of high-frequency noise.Lastly,time series can be regarded as a kind of data streaming whose distribu-tion would be varied along the time.Therefore,how to handle the drift problem is also a challenge.To deal with these problems,we propose two time series forecasting methods based on Recurrent Neural Networks and we have applied them to several real-world applications.The contributions are as follows:· We propose a GRU neural network based forecasting method combined with Multi-Lag ensemble and time series decomposition.We adopt a GRU network as the base model,which is specially designed for the sequence modeling so that it is suitable for the relevant data points in time series.In addition,we pro-pose a Multi-Lag ensemble learning strategy to cope with the window size and drift problems.It makes fuMl use of the advantage that GRU network can handle variable-length sequences.Finally,we introduce Filter Time Series Decompo-sition method.It can separate high-frequency noise from original time series to mitigate the problem of the high sampling rate.· Based on the above research,we propose a new RNN model,named Time Step Residual Recurrent Neural Network(TSR-RNN),which focuses more on the efficiency of training and testing processes and relevance of data points in time series.We add residual connections between different time steps,which can mitigate the gradient vanishing problem in the training process.Meanwhile,TSR-RNN is also able to learn the incremental patterns between time steps and owns better interpretability.Compared with GRU or LSTM networks,TSR-RNN can rival both of them in the aspect of efficiency and accuracy without extra training parameters.· Based on both of forecasting methods mentioned above,we implement them in two real-world applications.The first one is hardware resource forecasting of the large-scale cloud platform.The other one is Inventory Management and Forecasting Platform based on our algorithms.Through the experiments,we confirm that our proposed methods can handle the problems mentioned above very effectively.Compared with several commonly-used forecasting models,our methods achieve much better results than them.In practice,our forecasting methods also prove to be very accurate and robust.