Research Of Complex Time Series Forecasting Models Based On Deep Learning

Objective:Time series data in medicine often have the characteristics:complex data structure,huge data volume,high dimensions,high nonlinearity and lots of noise information.While traditional time series methods have been no longer appropriate for analysis and prediction of such complex time series,the emergence of deep learning methods,such as recurrent neural network(RNN),has provided new ideas for complex non-linear and non-stationary time series forecasting.This research proposed two new hybrid models based on RNN for time series prediction,in order to provide methodological support for further improving the prediction accuracy of such complex non-linear and non-stationary time series.Methods:(1)With two different strategies of neural network model optimization and time series decomposition,this study proposed two different complex time series forecasting methods based on RNN.To solve the problem of autocorrelation in time series data,we used gate recurrent unit(GRU)neural network as the basic model of our prediction methods,which is a special recurrent neural network.(2)For the difficult selection of neural network structure parameters,one of the swarm intelligence optimization algorithms,the three-dimensional chaotic fruit fly optimization algorithm(V3CFOA)was used to optimize GRU network parameters.V3CFOA combines the chaos algorithm to increase the diversity of the initial fruit fly population,and improves the two-dimensional search space to three-dimensional,which increases the search range and improves the convergence speed,thereby improving the algorithm's global optimization capabilities.Based on this method,this study will construct a V3CFOA-GRU hybrid model,and compare with other optimization algorithms and the GRU model through the case analysis,to verify the effectiveness of the hybrid prediction model.(3)Considering the characteristics of high noise and non-stationarity in complex time series,and starting from the perspective of signal decomposition,this study used the modified ensemble empirical mode decomposition(MEEMD)to decompose the original complex time series into relatively simple signals(subsequences)with different frequencies,and then GRU neural networks were used for those subsequences forecasting,to achieve the purpose of improving prediction accuracy.The MEEMD method decomposes sequences adaptively based on the time-domain local characteristics of the sequence.By adding paired white noise and setting permutation entropy detection,it solves problems such as mode mixing and large reconstruction errors,thus it can simply and efficiently decompose time series.Therefore,a GRU prediction model based on MEEMD will be constructed and compared with other decomposition methods and the GRU model through the case analysis,to verify the effectiveness of this hybrid model.(4)The case analysis data(Beijing PM2.5 dataset)is collected from the PM2.5pollutant concentration and meteorological data recorded every hour in 5 years by the US Embassy Monitoring Station in Beijing.The sequence length is 43824 hours.It has eight variables:PM2.5 concentration,dew point,temperature,air pressure,wind direction,wind speed,and accumulated hours of rain and snow.This study uses 8 variables in the past period to predict PM2.5 concentration in the next hour.The training set and test set are divided into 3:1.(5)In this study,the evaluation indicators of the prediction models are:root mean square error(RMSE),mean absolute error(MAE),symmetric mean absolute percentage error(SMAPE,%),and determination coefficient(R~2).The smaller the value of the first three error indicators,the better the prediction effect of the model;the closer the value of the determination coefficient is to 1,the better the model's fitting effect.Result:(1)A V3CFOA-GRU hybrid prediction model is constructed.The V3CFOA method is used to optimize the hyperparameters(lag time window and number of the hidden layer units)of the GRU.In the V3CFOA,RMSE of the test set is set as the fitness function,and it tries to find the optimal hyperparameters that minimize the RMSE.Then the optimal hyperparameters are used to build a GRU model for training and prediction,at last the prediction effect of the model will be evaluated.(2)The validity of the V3CFOA-GRU hybrid prediction model was verified.Firstly,compared with standard RNN and LSTM model at the same hyperparameter settings(set the lag time window as 12 hours,and the number of hidden layer units as 80),we got that,while the RMSE,MAE,SMAPE and R~2 of the RNN and LSTM were 24.24,14.22,30.63%,0.92,and 23.54,12.84,21.99%,0.93.respectively.the same indicators of the GRU were 23.29,12.50,20.08%,and 0.93,respectively.This shows the GRU model has a better effect of prediction than LSTM and RNN models.Besides,compared with FOA-GRU,PSO-GRU and GRU,it shows that,FOA-GRU,PSO-GRU both have better prediction results than GRU,as the RMSE,MAE,SMAPE and R~2 of FOA-GRU and PSO-GRU were 22.13,11.78,20.05%,0.93,and 22.74,12.07,19.91%,0.93,respectively.What's more,with the RMSE,MAE,SMAPE and R~2 as 21.27,11.32,19.47%and 0.93,V3CFOA-GRU has the best prediction effect,and in this model,the lag time window and number of hidden layer units are 8 and 72,respectively.(3)A MEEMD-GRU prediction model is constructed.The MEEMD is used to decompose the original sequence into some intrinsic mode functions(IMFs)and a residual component.Then the GRU model is used to train and predict each component separately,and the prediction results of each component are added to get the final prediction result.(4)The validity of the MEEMD-GRU hybrid prediction model was verified.At first,the decomposition results of EMD and MEEMD were compared.EMD decomposition obtained 19 components,and MEEMD obtained 16 components.The change charts of the residual components of the two decomposition methods both show an increasing trend of PM2.5 concentration during 5 years.Then,the prediction results of MEEMD-GRU,EMD-GRU and GRU models are compared and analyzed.While the RMSE,MAE,SMAPE and R~2 of the GRU were 23.26,12.34,19.87%,and 0.93,the same indicators of the EMD-GRU and MEEMD-GRU were15.78,9.66,21.78%,and 0.97,and 14.73,8.89,19.94%,and 0.97,respectively.It shows that the MEEMD-GRU model has the best prediction effect.(5)Compare the prediction results of MEEMD-GRU and V3CFOA-GRU:It shows that,the MEEMD-GRU model has better prediction effect than V3CFOA-GRU model.Conclusions:Compared with purely using GRU neural network and other hybrid models for time series forecasting,the V3CFOA-GRU and the MEEMD-GRU hybrid models both can get higher prediction accuracy.In addition,the GRU models combined with the sequence decomposition methods perform better than the GRU models combined with the parameter optimization methods,and this tells us that,to do some data processing on the complex time series before the analysis,can reduce the complexity of the series,which may be very helpful for further improving its prediction accuracy.When predicting of complex,high-noise,non-linear,non-stationary time series,both hybrid models can get high prediction accuracy.Therefore,we can use these hybrid models,according to the specific application scenario and the actual data quality.