การเปรียบเทียบอัลกอริธึม SES, DES, TES, ARIMA และ LSTM เพื่อการทำนายผลจากข้อมูลอนุกรมเวลาแบบสั้น

บทคัดย่อ

A time series is a sequence of observations which are collected and ordered over a period. The observations can be collected for hours, days, months, or years. There has been an assumption that the dependence of successive observations in time series is probable to exist. By means of time series analysis, this dependence is examined to discover a pattern utilised to prepare a prediction, and the prediction is one of various common objectives which can be achieved using time series analysis. Fortunately, time series analysis has been a topic of interest among academics for many decades and quite different algorithms for the prediction task have already been at our disposal. These algorithms are either well-established statistical ones or based on approaches of machine learning. In this paper, an experimental comparison of five different algorithms (simple exponential smoothing (SES), double exponential smoothing (DES), triple exponential smoothing (TES), autoregressive integrated moving average (ARIMA), and long short-term memory (LSTM)) is carried out using 10 examples of short time series to examine the performance of the selected algorithms over the task of time series prediction. The experimental results reveal that of 10 time series, the application of TES, and LSTM provides the best results in four time series each, while SES, and ARIMA outperform others in one time series each. It is more likely that LSTM algorithm could be used for the prediction task of time series analysis.