Research And Implementation Of Multivariate Time-series Prediction Model

Multivariate time series data are widely distributed in daily life,and multivariate time series prediction is becoming more and more important,which has been continuously researched in various fields such as transportation,finance and energy.In this context,this paper designs and implements a new multivariable time series prediction model and applies it to the actual metro passenger flow prediction data set.In the real world,time series data rarely exist alone,but usually appear simultaneously with multivariable sequence data,which interact with each other and have intricate relationships.In addition to the interrelationship between sequences,different periodic patterns within the sequence and the long-and short-term dependencies are also important conditions to achieve excellent prediction results.Most of the existing models are unable to cope with complex data in real scenarios,or are difficult to migrate from one scenario to another,resulting in poor generalization performance.Therefore,in addition to the traditional convolutional and recurrent neural networks,the time series clustering algorithm and self-attention mechanism are also integrated in this paper.The characteristics of multivariate sequence data and the disadvantages of existing models are comprehensively considered.A deep neural network model is established for multivariate sequence prediction.The model solves the problem of capturing the interrelation between sequences and multiple periodic patterns,and achieves good results in both the public and actual data set.The model in this paper is composed of data resolver,short-term prediction part,long-term prediction part and output layer.In the data resolver,the model divides the preprocessed data into two parts:long-term input and short-term input.The short-term prediction part adopts convolutional and recurrent neural network to learn various periodic information and short-term dependence in the sequence.In the long-term prediction part,time series clustering algorithm and self-attention mechanism are used to capture correlation and long-term dependence between sequences.The output layer fuses the representation of the two parts to produce the final predicted result.In this paper,the advantages and universality of the proposed model are illustrated by the comparative experiments on multiple horizons of multiple data sets,and the results of the actual metro passenger flow data set are also better than the existing model,with strong generalization performance.