Improving long-term multivariate time series forecasting with a seasonal-trend decomposition-based 2-dimensional...

Area of Science:

• Time Series Analysis
• Machine Learning
• Deep Learning

Background:

• Transformer-based models often underperform simple linear models in long-term multivariate time series forecasting.
• Existing methods struggle to capture complex interdependencies and temporal features like seasonality and trends.

Purpose of the Study:

• To propose a novel model, STL-2DTCDN, that addresses limitations in current long-term multivariate time series forecasting.
• To improve the accuracy of forecasting by effectively utilizing temporal features and inter-series dependencies.

Main Methods:

• Incorporation of Seasonal-Trend decomposition based on LOESS (STL) to extract trend and seasonal components.
• Design of a 2-Dimensional Temporal Convolution Dense Network (2DTCDN) to model complex interdependencies among multivariate time series.
• Evaluation of the proposed STL-2DTCDN model on six diverse datasets.

Main Results:

• The STL-2DTCDN model demonstrated superior performance compared to existing methods in long-term multivariate time series forecasting.
• The model successfully leveraged both seasonal-trend features and complex interdependencies for improved accuracy.

Conclusions:

• The proposed STL-2DTCDN offers a significant advancement in long-term multivariate time series forecasting.
• STL-2DTCDN provides a robust framework for accurately forecasting complex time series data by integrating decomposition techniques with advanced deep learning architectures.