Research And Application Of Neural Network In Nonstationary Time Series Prediction

The advancement of network information technology has led to the rapid development of the era of big data,and many time series data are recorded in every moment of life.Faced with the changes of random and complex time series data in reality,how to extract and analyze the deep non-stationary time series information embedded in the data,and study the historical data through the time sequence to predict the trend in the future period is gradually becoming a new challenge in the field of time series forecasting.In recent years,neural networks have been widely used as a novel modelling solution for non-stationary time series forecasting.However,due to the volatility of non-stationary time series,modal confounding and noise interference are prone to occur when dealing with such data,resulting in incomplete extraction of sequence information,and most neural networks,despite their strong non-linear learning capability,can suffer from overfitting and slow convergence during the prediction process,thus affecting the prediction results.Based on previous research,this paper investigates and innovates the traditional decomposition method for processing non-stationary time series,and optimize the internal structure of the neural network,while incorporating an improved optimization-seeking algorithm to find the optimal parameters for prediction of a specific data set,to make more accurate prediction of non-stationary time series through a hybrid neural network model.Firstly,this paper proposes an EVMD-TBPNN-CLSTM algorithm based on Long Short-Term Memory(LSTM)for non-stationary time series prediction.Energy Variational Mode Decomposition(EVMD)is used for the decomposition of non-stationary time series to determine the optimal number of decomposition modes suitable for a particular data set;Secondly,based on the different modal characteristics of the EVMD decomposed series,an improved African Vulture Optimization Algorithm is used to optimize the Back Propagation Neural Network for the smoother series,and a hybrid neural network model for prediction of the decomposed high-frequency non-smooth sequences,which is reconstructed and summed to obtain the point prediction values.Finally,based on the errors generated during the training of the calibration data for EVMD,the mean absolute percentage interval estimate forecasts are constructed on the basis of the point forecast model,thereby improving the interval coverage accuracy of the non-stationary time series forecasts.Secondly,this paper proposes a non-smooth time series prediction algorithm SCEVMD-GELU-TCN based on Temporal Convolutional Network(TCN),which introduces a dual approach of sample entropy and correlation coefficient to determine different frequency modal components on the basis of EVMD,and jointly adaptive wavelet threshold function The method performs a secondary noise reduction process on the mixed components of the residual components,while discarding the less influential noise components to achieve the purpose of attenuating the component noise interference.At the same time,in order to better combine the non-linear characteristics of the activation function inside the neural network with the stochastic regulator,this paper improves the activation function in TCN as well as the residual structure to retain the complete information of the sequence,so as to improve the network prediction performance.Finally,wind power time series prediction is the main application scenario for non-stationary time series prediction algorithms.In this paper,we try to apply two types of neural network-based non-stationary time series prediction algorithms to the wind power and wind load prediction problems in wind power prediction.In this section,the theory related to the wind power time series prediction problem is introduced in detail,and the proposed EVMD-TBPNN-CLSTM and SCEVMD-GELU-TCN are applied to the prediction of KDD Cup 2022 wind power dataset and Elia wind load dataset,Respectively,according to the applicability of the algorithms,so as to verify the two types of algorithms in the non-stationary time series prediction scenario The effectiveness of the two algorithms in non-stationary time series forecasting scenarios was verified.