Research On Combined Prediction Model Of Time Series Based On Deep Learning

With the development of the era of big data,the amount and types of data are constantly increasing.Among them,time series data is an important data form,which is widely and diversely distributed,and its research value is becoming increasingly prominent.Time series prediction is one of the research branches in the area of signal processing.By means of time series analysis and prediction,scientific community uses various algorithms to model and process time series,mining its internal rules and predicting its future values.A high-precision prediction model can help researchers make better decisions,thus reducing risks and increasing benefits,and has a high application value in many scenarios.In recent years,deep learning technology has attracted extensive attentions due to its excellent feature extraction ability and efficient generalization performance,providing a new approach for time series prediction.However,the time series data appears with non-stationary,nonlinear and multi-dimensional characteristics,it is difficult for a single deep learning framework to effectively extract the deep features of time series and identify the potential patterns between data.Therefore,based on the research of various current deep learning models,this paper proposes the following two combined prediction models:(1)A combined time series prediction model based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN)and Temporal ConvolutionalNetwork(TCN),i.e.,a prediction model of CEEMDAN-TCN.First,in view of the nonlinear and non-stationary characteristics of time series data,by means of CEEMDAN,the time sequence is decomposed into several sub-sequences with different frequencies to reduce the complexity of the original time series.Then,TCN is used to model each sub-sequence.Finally,all the outputs of the TCN models are summed up to form the final prediction results.The experimental results indicate that compared with the Long Short-Term Memory model and other hybrid models,the proposed CEEMDAN-TCN model shows a better performance in both univariate and multivariate prediction tasks.(2)A time series imaging prediction model based on Densely Connected ConvolutionalNetworks(DenseNet)and Temporal ConvolutionalNetwork(TCN),i.e.,a prediction model of DenseNet-TCN.Firstly,the time series data is encoded into several single-channel images by Recurrence Plot,Gram Angle Field and Markov Transition Field respectively.The resulting single-channel images are then stacked to generate several multi-channel images.Finally,features are extracted by the DenseNet,and the TCN model outputs the final prediction results.Experiments show that this method can further extract the time correlation information and pattern information,which is suitable for long-term prediction tasks.