Joint BCAM-UPV/EHU Data Science and Artificial Intelligence seminar: Forecasting Sequences of Time Series: An Experimental Comparison

LOCATION: UPV/EHU Donosti, Faculty of Computer Science, room 3.1 and online

Link to the session here

Abstract
Time series (TS) forecasting has become a broad and active research area in the past decades and a large portfolio of methods have been developed in order to perform forecasts for different application domains. To the best of our knowledge, no work has addressed the problem of forecasting when the data are sequences of complete time series. This type of data appears naturally in industrial contexts, for example, when periodical tests are performed to the machines to determine their condition, and each test is represented by a TS. In the present work, we formulate the problem of forecasting in sequences of time series. The objective in this problem is to forecast the whole time series based on the sequence of past observations. Then, we perform an experimental analysis to compare two novel approaches that can be used to solve this task: (i) unidimensional and (ii) multidimensional approaches. The experimental work to evaluate the proposals is performed using a publicly available dataset obtained from commercial lithium iron phosphate/graphite cells cycled under fast-charging conditions. The results indicate that considering the correlations between the values in the TS of the sequence improves the forecasts.