Variational Mode Decomposition dan Linear Embedding adalah yang Anda Perlukan untuk Prediksi Data Deret Waktu

Abstract

One of the challenges in time-series forecasting is data volatility. Volatilities in data may lead to inaccurate forecasting results. Applying variational mode decomposition (VMD) are one of many approaches to reduce volatilites in data. Considering Long-term Time Series Forecasting Linear (LTSF-Linear) performance against Transformer based model, linear based model were used together with VMD to develop a forecasting model with low error rates. This research used 13 different datasets, including ETTm2, WindTurbine, M4 Dataset, and 10 air quality datasets from different cities in Southeast Asia. The VMD strategy was utilized to break down the information into numerous modes or data components. Root Mean Squared Error (RMSE) were used as metric evaluation method. RMSE values attained from forecasting models that utilized VMD are compared with those without VMD. Furthermore, we also compared linear-based models with some well-known neural network models such as LSTM, BLSTM, and RNN by comparing their RMSE values. The results underscore VMD’s superrior performace, nearly all models experienced a significant reduction in RMSE values after applying the VMD decomposition strategy. Linear + VMD model achieved the lowest average RMSE value in univariate forecasting at 0.619. In multivariate forecasting, DLinear + VMD model emphasized its superiority by achieving the lowest RMSE values on all multivariate datasets with average RMSE of 0.019. These results demonstrate the superiority of LTSF-Linear models combined with VMD on time series forecasting.