Time Series Prediction And Classification Based On Deep Learning

In recent years,with the rapid development of the Internet of Things and 5G technology,time series data has shown a growing trend.These time series data contain a lot of hidden information,and mining and analyzing these hidden information is of great significance in the fields of finance,medical treatment and transportation.However,the existing time series prediction and classification models have failed to fully consider the characteristics of time series data(ie,the correlation and long-term dependence between sequence variables).Traditional machine learning algorithms extract data features through manual design rules,while deep learning learns abstract representations of data through multiple processing layers,which not only saves the steps of manually extracting features but also greatly improves the generalization performance of the model.Therefore,this paper uses deep learning to study the two key issues of time series data prediction and classification.The main work contents are as follows:(1)Investigate the traditional prediction and classification methods of time series data,and at the same time investigate the relevant theoretical knowledge of deep learning,including convolutional neural networks,recurrent neural networks and attention mechanisms,and focus on the study of time series prediction and classification methods based on deep learning.(2)In view of the problem of the lack of considering the correlation between sequence variables and long-term dependencies in the time series prediction model CRNN(Convolutional Recurrent Neural Network),this paper proposes an improved model SACGRNN(Self Attention and Convolution based Gate Recurrent Neural Network)for CRNN.The model first learns the time-invariant features of each time series through a convolution kernel,then learns the correlation between time series variables by introducing a self-attention mechanism,and simultaneously captures the time dependence of time series data and time series data scale changes through GRU and Auto Regression.This paper proves that the prediction effect of the SAC-GRNN model on the multivariate time series data set is better than the comparison model by setting different prediction step experiments.(3)Traditional time series classification methods extract sequence features by designing different rules.When the scene changes,the fixed feature extraction rules will fail.In order to effectively classify different time series,this paper proposes a dual-channel convolution-gated neural network time series classification model DCC-GNN(Dual-channel convolution-gated neural network).The model not only extracts high-level feature representations of different time series through dual-channel convolution,but also uses the gating unit to model the time dependence.In this paper,we compare experiments with different neural network models on 85 data sets of UCR warehouse,and the experimental results show that the DCC-GNN model has a good classification effect.