DDformer: Decomposition and Dimension Transformer for Multivariate Time Series Forecasting
Shotaro Kawano, T. Kawahara · 2024
Recently, the large amounts of time series data generated by IoT devices are used for forecasting. Various multivariate time series forecasting models have been developed using deep learning models. Among them, Transformer-based models, which can extract long-term dependencies within sequences, have attracted significant attention. However, it is necessary for Transformers to effectively capture dependencies between multiple time series data. Additionally, simplifying the structure is required for implementation on IoT devices, and there is also a need to develop models that mitigate the impact of noise present in time series data. In this paper, we propose a Transformer-based model called DDformer to address these challenges. DDformer is designed to effectively capture both temporal and spatial dependencies in time series data. It decomposes inputs into trend and seasonal components using decomposition layers and enhances the features of each time step and variable with dimension expansion/reduction layers. When validated on energy, financial, and weather datasets, DDformer reduced prediction error by up to 45.9% compared to the state-of-the-art model (FEDformer).