Time Series Prediction And Classification Method Based On Deep Learning

Time series is a kind of high-dimensional data that exists widely in various fields of reality.In the research activities related to time series,time series prediction and time series classification are two research priorities.The traditional time series prediction method only analyzes the time series from the time dimension,ignoring the influence of external influence factors on the time series.The traditional time series classification method relies too much on the similarity measure between sequences,ignoring the inherent law of the time series itself.In this paper,the deep learning method is used to solve the problems existing in traditional time series prediction and traditional time classification methods.The main research contents are as follows:(1)This paper summarizes the research status of time series prediction and classification,and focuses on the research status of time series prediction and classification based on deep learning.The theories and methods related to deep learning are investigated,including back-propagation neural networks,convolutional neural networks,recurrent neural networks,long-short-term memory networks,and attention mechanisms.(2)Aiming at the problem that the time series prediction model DA-RNN(Dual-stage Attention based Recurrent Neural Network)can't model the correlation between input features and predicted features and the correlation between input features,this paper proposes a DAFDC-RNN(Dual-stage Attention and Full Dimension Convolution based Recurrent Neural Network)model.DAFDC-RNN corrects the part of DA-RNN's problem definition that is inconsistent with the actual application scenario,and introduces the target attention mechanism to learn the correlation between the input feature and the predicted feature,and introduces a full dimension convolution mechanism to learn the correlation between input features,and introduces a temporal mechanism to learn the long-term temporal dependencies of time series.The experimental part firstly determines the hyperparameter of the model,and then verifies the structure and components of the model.Finally,the comparison experiments show that the proposed DAFDC-RNN model has better prediction effect on the large feature quantity dataset than the DA-RNN model.(3)DenseNet is a neural network model based on dense connection.This paper proposes a time series classification model based on DenseNet.In order to verify the classification performance of the model,this paper conducts comparative experiments on 85 data sets in the UCR warehouse,and uses the data visualization technology CAM to explain the decision process of the classification model.The experimental results show that DenseNet's classification performance is better than the frontier time series classification model,including Residual Network(ResNet),Full Convolutional Neural Network(FCN),Multi-scale Convolutional Neural Network(MCNN)and Multi Layer Perceptron(MLP).).