Hybrid Prediction Of Time Series Based On Latent Feature Extraction

Since time series prediction has a wide application in various engineering and industrial practices,studies upon it remain hot in past decades.Fruitful achievements have been obtained,especially for the hybrid structure-based prediction models.In such a structure,the prediction performance can be effectively improved through the ensemble of multiple prediction models to get the changing features of the time series.Previous studies have shown that prediction models' performance heavily relies on the good representation of processed data.A high-level representation can effectively mine the latent features of data.However,no existing research has ever considered extracting those valuable latent features of time series,not to say those latent-feature based hybrid prediction methods.Motivated by these,the hybrid prediction strategy based on latent feature extraction will be further studied in this paper,and two main research contents are as follows:(1)Static Hybrid Prediction of Time Series Based on Latent Feature Extraction with fix Sliding Window: Inspired by the existing three types of hybrid prediction structures,the general framework of the latent feature extraction-based hybrid prediction model for the time series is firstly presented.Then,a specific algorithm by combining the prediction strategy and latent feature extraction within a fix sliding window is proposed.A Long Short Term Memory network is used to predict the time series within a fixed-length sliding window and extract the latent features of the series in each window.With these latent features,the hybrid strategy of linear weighted multiple prediction models is adopted to construct the static parallel combination prediction model.The presented algorithm is applied to four typical time series data sets,and the experimental results demonstrate that the algorithm can effectively improve the prediction accuracy.(2)Dynamic Hybrid Prediction of Time Series Based on Latent Feature Extraction: The outcome of research(1)can hardly be applied to those scenarios where time series change dynamically since the length of sliding window and the parallel combination weights remain unchanged.Given this,an improved hybrid prediction algorithm with dynamically changing the length of sliding window and the parallel combination weights is further proposed.First,according to the changeable prediction accuracy of the hybrid model,the strategy of adapting the sliding window length and latent feature extraction update is given.Then,as for the parallel combination prediction module,a dynamic adjustment strategy of the combination weights based on the prediction accuracy is proposed.Based on the first two improvements,the hybrid prediction of time series is realized.The application of the algorithm in typical data sets demonstrates that it can further improve the prediction accuracy of time series and effectively promote the performance of the prediction model.Aiming to improve the technical defects existing in hybrid prediction models for time series,a hybrid prediction architecture for time series based on the latent feature extraction and multi-model parallel combination is firstly proposed.Then,for the latent feature extraction module and the multi-model parallel combination prediction module,a fix sliding window and static combination weights,as well as the variable sliding window and dynamic combination weights strategies are considered,respectively,and the experimental results show the effectiveness of the proposed algorithms.This thesis contains 22 figures,10 tables and 98 references.