Time Series Forecasting Based On Deep Learning

Time series analysis is a science of data processing and modeling on time series data at a certain time interval.It plays a vital role in quite a few fields,such as finance,medical care,meteorological observations,censuses,web traffic,and transportation.As one of the core tasks of time series analysis,time series forecasting infers and extends according to the development and trend of time series,and subsequently predicts the state of next moments.With the generation of massive data,the time series forecasting task changes from single sequence to multi-sequences,from low-dimensional time series features to high-dimensional complex time series features.The time series forecasting model also changes from the classical autoregressive model to the complex nonlinear model in accordance with the change of data features.For instance,deep learning models such as convolution neural networks and recurrent neural networks were adopted.The sequence models in deep learning are closely related to the multi-step prediction issues in time series.They are both involved with modeling based on known sequences and outputting new sequences of unequal length.The encoding-decoding framework is one of the effective ways to process the sequence model.Different network models can be used in both encoding and decoding stages.In addition to the recurrent neural network GRU in this paper,the time convolution network is used to extend the input data span and increase the historical information covered by each neuron node of GRU at the encoding stage.Based on the "encoding-decoding" framework,this paper also introduces multi-attention mechanisms to lead the model to focus on local features.One is to add self-attention layers to generate self-attention features based on the relevance of input vectors.The other is to construct the intermediate vectors corresponding to different decoding steps in the use of sliding window,so as to increase the model flexibility,reduce parameters and speed up the model operation.A time series forecasting model based on these two types of attention mechanisms was proposed in this paper.Finally,aiming at the multi-step forecasting of multi-sequence,web traffic data was used for experimental research.Utilizing TensorFlow-GPU to realize deep models,various deep models based on "encoding-decoding" framework can be compared to verify the algorithm effect.The model based on time convolution network and two-stage attention mechanisms turns out to be more effective.