Deep Learning Model Based On The Combination Of Long And Short Memory And Convolution And Its Application

Time series data are easy to acquire and widely exist,which makes the research on time series more and more important.Therefore,more and more researches on time series classification task and time series prediction task have been proposed by relevant researchers.However,the existing methods mainly have the following disadvantages: First,many existing methods need to manually measure the time series features,but the manual measurement has a strong subjective dependence on the criteria and thresholds for the division.And the second,due to the existence of outliers and noise in the time series,the quality of the extracted features in the existing methods is hard to measure.Third,most of the existing methods only consider the local features in the time series,they can not capture the long-term dependencies that exist in time series data.And there is no one kind of distance measurement method suitable for all fields.In this paper,we take the financial markets and the ocean data for the research background,the stock price and the surface temperature of the ocean as the research object,the main research contents are as follows.Firstly,considering the existing problems in time series classification,this paper proposes a time series classification model(LCNN)which combines Long Short Term Memory Network(LSTM)and Convolutional Neural Network(CNN).The model mainly adopts multi-branch network structure,LSTM is used to process and predict the important events with relatively long intervals and long delays in the time series,and the temporal features in the time series and the long-term dependencies in the time series are extracted.Then,the convolution kernel of multi-granularity in convolutional neural network is used to convolute the time series to mine the local and deep features of the time series.Finally,experiments were performed on 15 datasets in the UCR open dataset to compare LCNN with 16 existing time-series classification methods.Secondly,a comprehensive evaluation model(CEM)based on traditional methods is proposed.The model is mainly divided into three parts: the traditional statistical methods are used to analyze long memory,complexity and time-series asynchrony in time series.The long memory of the time series is evaluated by using R/S method and the long memory of the time series is quantitatively measured by using the Hurst index.Then the time series to be studied as a system to explore the complexity of the system,that is,the stability of the time series.Finally,we study the asynchronism between time series.Thirdly,applying LCNN model and CEM model to the stock market and sea surface temperature(SST),we first use CEM model to study the long memory,complexity and asynchrony of stock closing price and sea surface temperature data.Then,based on the LCNN model,the trend of stock price and the sea surface temperature are predicted.