Time Series Prediction Of Neural Network Model Based On Variational Modal Decomposition

Complex systems widely exist in many fields such as meteorology,hydrology,finance and so on,and their internal dynamic evolution rules often present complex chaotic characteristics.At present,people use the collected time series of complex systems to study them.Due to the non-stationary characteristics of the collected time series,the simple-structured prediction model is difficult to meet the actual prediction accuracy requirements.Therefore,this paper considers the method of combining signal decomposition method and neural network model to predict time series.The signal decomposition method decomposes the time series into multiple frequency sub-signals,and modeling and predicting the sub-signals by constructing a neural network model is beneficial to improve the prediction accuracy.At present,signal decomposition methods mainly include empirical mode decomposition methods and variational mode decomposition methods.The empirical mode decomposition method has the disadvantage of modal aliasing,which leads to the inaccuracy of the decomposition subsequence.Variational modal decomposition method is a non-recursive decomposition method to avoid this shortcoming.In this paper,a combined neural network model based on variational mode decomposition is proposed.The original time series is decomposed into a predetermined number of subsequences by variational mode decomposition.The subsequences are extracted by using the deep belief network,and finally the most predicted results are output by the rebound state network model.This combination model avoids the tedious process of building models for subsequences and ensures the prediction accuracy.In addition,in order to simplify the construction of subsequence prediction model,this paper proposes the variational mode decomposition method and the improved echo state network prediction model.The permutation entropy is used to estimate the complexity of the subsequence and the original sequence,which ensures that the decomposed subsequence has lower complexity than the original sequence,and the subsequence has a stable trend of change.The improved echo state network model is used to predict the subsequence,which has strong generalization ability and avoids the tedious process of building prediction models for subsequence.The simulation results of the real data set show that the model can effectively improve the prediction accuracy.The actual time series has strong non-linear and non-stationary characteristics,so it is difficult to make long-term prediction of time series.In this paper,a new method of variational mode decomposition and a combined model of double hidden layer recurrent neural network are proposed.The original sequences are presented as multi-dimensional information subsequences by using the variational mode decomposition method.The decomposed multi-dimensional subsequence can show the effective characteristics of the complex system,which is helpful for the neural network model to predict the time series.In order to fully extract the feature information of the decomposed subsequence and its internal long-term evolution rules,an information complementary neural network model is proposed.The neural network model has a dual reservoirs structure.The original information and decomposition information are extracted by using the first reservoir,and the extracted information is combined with the decomposition subsequence to use the second reservoir for prediction.The simulation results of London's NO₂ data set show that the short-term prediction and long-term prediction show a high prediction accuracy.