Multiple Convolutional Neural Networks For Multivariate Time Series Prediction

Multivariable time series prediction has become an important part of machine learning(including deep learning).Because with the in-depth study of this issue on human production and life has gradually produced a far-reaching positive impact.However,for the researchers in this field,the task of time series prediction still faces a very difficult challenge,because the prediction target is affected by many complex factors.For example,when forecasting the traffic flow on the highway or the generation of solar power station,the change of meteorological conditions will have a significant impact on the prediction target,and the existence of such uncertainty will lead to the deviation of the prediction results,or even cause serious problems.Especially for time series with strict periodicity,if periodic timing information can be extracted to the greatest extent from historical observation data,it means that the accuracy of prediction will be greatly improved.At present,the time series prediction model based on RNN(Recurrent Neural Network)has made great progress in different time series prediction application tasks.However,this type of prediction model is difficult to obtain the overall structure information of time series,resulting in the model can not effectively extract the periodicity characteristics of time series.This problem can be solved by the CNN(Convolutional Neural Networks)model.Therefore,this paper proposes a time series prediction model called "multiple CNN" to solve the time series prediction problem with strict periodicity.The working principle of multiple CNN is to extract the compactness information of the target time point and its nearest time point as well as its own periodicity information by analyzing the periodicity of the time series,and finally make a comprehensive prediction based on the characteristics of these three parts.At the same time,the model is highly flexible,allowing users to freely adjust the cycle range set in the model according to the characteristics of their data sets.In view of the shortcomings of the aforementioned RNN model in time series prediction and our research on time series prediction with multiple CNN model,this paper mainly carries out the following work:1)Based on the extraction and analysis of periodic features of time series,this paper proposes the three-part feature fusion method for time series prediction.The three features correspond to the influence of the nearest time period on the predicted time point and the characteristic information of the long period and short period of the predicted time point respectively,and the formal definition and description of the three features and fusion process are given.2)At present,as the most popular third-party open source library,Tensorflow has been widely used in different research fields of machine learning.Therefore,after analyzing the characteristics of time series,this paper uses CNN model to conduct modeling operations for the three parts,and finally conducts experimental analysis in the Tensorflow environment.This paper uses two large real-world data sets to verify the validity of the proposed multiple CNN model.At the same time,in order to ensure the fairness of the experiment,the multiple CNN model is used to compare with other time series models under the same environment and super parameter setting.Finally,these empirical comparative experiments prove that the proposed multiple CNN time series model has a strong advantage in prediction accuracy.