Research On Correlated Multiple Time Series Prediction Based On Deep Learning

Time series data is data collected at different times and used to describe how the phenomenon changes over time.This type of data reflects the change state or degree of a certain thing,phenomenon,etc.over time.Time series data is widely available in real life,such as transaction data and economic statistics in the financial field,user browsing and purchase data in the consumer e-commerce field,signal records of medical devices in the medical field,and weather recorded by weather monitoring stations These time series data such as indicator data are very valuable data resources in the corresponding fields.The accurate and effective analysis and utilization of these data can help reduce labor costs,improve production efficiency,and increase economic benefits.Time series prediction is usually a non-linear fitting problem that relies on interrelated multiple factors.Our predict of correlated multivariate time series should also take into account the time dependence and variable dependence within the series and the spatial dependence between the series.Traditional time series models such as autoregressive moving average and differential integrated moving average autoregressive are difficult to effectively model complex nonlinear time series data;machine learning methods such as multiple linear regression and neural networks cannot capture well The long-term dependence in the time series can only use fixed-length data sequences to predict,which is limited by the model capacity,and it is difficult to achieve the best prediction effect;the long-term short-term memory model(LSTM)is based on the existence of "memory" on the cyclic neural network.The “body” model can remember information for a long time,effectively solve long-term dependence problems,and as a deep neural network,it has super fitting capabilities and has natural advantages in the processing of complex nonlinear time series data.Therefore,This paper proposes a universal time series prediction model CALSTM(Convolution and Attention based on Long Short Term Memory)based on LSTM.The convolutional long-term and short-term memory model can effectively extract the variable-dependent features and time-dependent features within the sequence and the spacedependent features between sequences,so that the algorithm considers information more comprehensively and the operating structure and rules are clearer;attention mechanism Weighting the hidden layer state of all time steps,focusing on the more important hidden layer state information at all time steps,enhancing the interpretability of the model,and making the prediction effect for longer time series more obvious.In order to verify the validity and versatility of the model,this paper uses two related multivariate time series prediction examples of air quality prediction in the environmental field and taxi passenger demand prediction in the transportation field to normalize the data,deal with missing values and outliers,Feature coding,PCA dimensionality reduction,and selection of gradient descent methods,hyperparameter adjustment and other methods to optimize the model,use root mean square error,average absolute error as evaluation indicators,and draw training,test loss graphs,fitting graphs,and Multi-level attention mechanism model,LSTM model,multiple linear regression model,fully connected neural network model,differential integrated moving average autoregressive model five benchmark model prediction results are compared,the experimental results show that CA-LSTM has a better prediction effect.