|
Abstract
|
Accurate time series forecasting is crucial in various fields, including finance, economics, healthcare, transportation, and energy. Recently, deep learning methods have gained significant attention in time series forecasting. Despite their ability to model complex and nonlinear data patterns, their performance may be degraded when there is not sufficient input data. To address this issue, in this study, we propose a methodology that performs feature-based representation through time series featurization and adds the relevant features as an auxiliary input to deep learning models. Specifically, after conducting the featurization step, as another contribution, we develop a novel combined feature selection algorithm called JAYA-Genetic (JaGen) and employ it as the feature selection technique. The JaGen algorithm takes advantage of both the JAYA and Genetic algorithms to make balance between exploring and exploitation processes. To assess the efficacy of the suggested methodology, two deep learning models, a convolutional network and a multi-head attention mechanism are used. According to the findings of experiments carried out on ten Covid-19 datasets, two product demand time series datasets, and two public time series datasets, the derived models outperform their regular counterparts and current models in terms of performance metrics like symmetric mean absolute percentage error (SMAPE) and root mean square error (RMSE).
|