Experimental Comparison of Some Classical Distance Measures for Time Series Data in Simulation Model Validation
Xiaojun Yang, Zhongfu Xu, Haibo Ouyang, Xing Zhang · 2019 IEEE 8th Data Driven Control and Learning Systems Conference (DDCLS) · 2019
The comparison of simulation-generated time series with observed data is a basic problem in simulation model validation field. Many distance measure methods have been proposed or used by a number of modelers and users to validate simulation models for several decades. However, efforts to evaluate the performance of these methods are rarely published, especially on large datasets. Therefore, in this paper, the accuracies of some classical distance measures in simulation model validation are tested and evaluated on the latest UCR time series archive. Experimental results suggest that Theil's inequality coefficient and Manhattan distance outperform other distance measures, and elastic and normalized distance measures are generally effective to reflect the underlying dissimilarity of time series data. Furthermore, we also expect our work may promote quantitative comparisons by providing a common framework for experimental research in simulation model validation community.