부산대학교 도서관

For forecasting of energy related time series data (e.g. "load" or "generation"), many different kind of learning algorithms exist. The task of selecting an appropriate algorithm which is adequate for usage with a pre-given time series is not easy for non-data science experts. Since a trial-and-error approach for finding a suitable algorithm is tedious, computationally intensive and time-consuming, meta learning approach which describes how to design an appropriate methodology to constrain the search space for fully automatically finding the suitable learning algorithm is proposed. In the present paper, based on the assumption that good indicators describing certain characteristics of an e.g. energy time series dataset may provide useful insight into which forecasting algorithms are most suitable, a Descriptive Statistics Time-based Meta Features (DSTMF) description format for energy time series datasets is proposed to efficiently select an appropriate learning algorithm to perform energy forecasting on a time series dataset. In order to demonstrate the performance of the new methodology, experiments on datasets which are load time series from 60 institutional buildings are conducted. Based on a similarity-based clustering analysis, the potential of DSTMF's meta features for capturing deep characteristics of energy time series datasets is evaluated and compared to other state-of-the-art meta features. The experiments show very good results and outperform state-of-the-art meta features to enhance model selection in energy time series forecasting.