Research On Prediction Method Of Nonlinear Time Series Data With Complex Noise

Complex time series forecasting problems generally exist in complex systems such as intelligent transportation,weather forecasting,food safety,and financial economy.They are closely related to daily life and closely related to social development and economic development.It can help managers make better decisions,improve efficiency,and reduce losses with accurately grasping the fluctuation direction of the time series.The difficulty of prediction is increased due to the high degree of volatility and complexity of the time series obtained.It is still a challenge to predict the nonlinear time series with complex noise accurately.This paper starts with the study of nonlinear prediction methods for time series data to improve the accuracy of predicting its future development trend with Gated Recurrent Unit(GRU),Convolutional Neural Network(CNN),Empirical Mode Decomposition(EMD)and covariance Covariance Intersection(CI)fusion and other methods based on the feature extraction method of different patterns in complex time series.The main work of this article can be divided into the following three parts:(1)Because of the uncertainty of the number of EMD decomposition components,this paper proposes a component adaptive combination decomposition method based on CNN network-CEMD.This method first uses the EMD algorithm to decompose the wave components with different statistical characteristics in the time series,thereby reducing the influence of uncertain factors on the data and providing more useful data for subsequent processing.After that,a one-dimensional CNN network is used to identify and classify the components obtained by EMD decomposition.Finally,the components with similar wave pattern characteristics are reconstructed by adding,to get a fixed number of combined components.The combined components obtained after reconstruction are more suitable for the construction of prediction models and can guarantee the models' generalization ability.(2)In the construction of the combined component model,this paper analyzes the shortcomings of the currently widely used recurrent neural network and proposes a prediction unit-CEGRU based on the CEMD decomposition method,which uses the GRU deep learning network as The sub-prediction model models the combined components,which can make the model have better learning ability and prediction accuracy.(3)Aiming at the actual available dataset size with complex noise,this paper builds a deep hybrid prediction model based on CEGRU.It proposes a model SCEGRU for small-scale data set prediction and a model LCEGRU for large-scale data set prediction.The SCEGRU model first uses the random probability method to perform data enhancement processing and reorganization on the original data of the time series and then uses the proposed CEGRU prediction unit to model the reorganization sequence.The training of the LCEGRU model is continuously updated along the time axis with the time series data,and the model is saved every time the data is updated.The above two models will produce multiple component models during the training process.For this reason,the CI algorithm is introduced in the prediction stage to fuse multiple prediction results,and then the prediction output of the original time series is obtained.The deep hybrid prediction model studied in this paper can effectively solve the problem of nonlinear non-stationary time-series data prediction with complex noise.It also shows good prediction ability on multiple types of real data sets.