A granularity time series forecasting model combining three-way decision and trend information granularity

Abstract

In recent decades, significant advancements have been made in the field of time series data mining, leading to its widespread application in various domains. However, the existence of data correlation in time series data sets brings challenges to long-term prediction. One approach to address this issue is to transform the original time series into granular time series (GTS). Therefore, a prediction model based on GTS is proposed to meet this challenge in this study. Firstly, the improved L1-trend filtering is used to achieve the optimal segmentation of information particles. Then, a three-way decisions (TWD) model based on shape similarity is used to compress and aggregate information granules. Finally, a GTS prediction model based on LSTM neural network is established. The model effectively retains the trend information of the time series and overcomes the limitation that the existing models cannot adjust the granularity length of the original information. In addition, the proposed model is applied to several real datasets for sensitivity analysis and comparative analysis. The results show that the model has strong performance in long-term forecasting.