A Time Series Short-Term Prediction Method Based on a Multigranularity Event

Abstract

Time series forecasting with high accuracy is essential to many monitoring applications, such as traffic flow, power load, financial risk and agricultural monitoring. Considering the cost of computing resources, pattern recognition is still an important approach. However, in the segmentation step, most of the traditional methods use fixed-length segmentation. This causes poor performance in time series prediction. To overcome these problems, a short-term forecasting method for time series based on a multigranularity event, MGE-SP (multigranularity event-based short-term prediction), is proposed in this paper. First, a methodological framework to improve the prediction effect is established to guide the implementation steps. Next, three key methods in the framework are described in detail, including multigranularity event matching based on the LTF (latest time first) strategy, time series alignment using piecewise aggregate approximation based on the compression ratio and a short-term prediction model based on XGBoost. Then, the MGE-SP is tested with a practical dataset and an open dataset. The experimental results show that the proposed framework and methods result in a better effect in short-term time series prediction.

Keywords: time series, Short-term prediction, Multigranularity event, Alignment, Event matching