DiffMixer: A prediction model based on mixing different frequency features.

Area of Science:

• Data Science
• Machine Learning
• Signal Processing

Background:

• Time series forecasting is crucial in energy and network security.
• Existing Transformer and MLP models struggle with non-stationary real-world sequences.
• Traditional methods for non-stationarity either remove useful patterns or inadequately capture them.

Purpose of the Study:

• To propose DiffMixer, a novel method for analyzing and predicting different frequencies in non-stationary time series.
• To overcome the limitations of existing models in handling complex temporal dependencies and non-stationarity.
• To improve the accuracy and robustness of time series forecasting.

Main Methods:

• Variational Mode Decomposition (VMD) to extract frequency components.
• Multi-scale Decomposition (MsD) for optimized decomposition of downsampled sequences.
• Improved Star Aggregate-Redistribute (iSTAR), Frequency domain Processing Block (FPB), and Dual Dimension Fusion (DuDF) for inter-component analysis and fusion.

Main Results:

• DiffMixer demonstrated significant reductions in forecasting errors.
• Mean Squared Error (MSE) decreased by 24.5%.
• Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Symmetric Mean Absolute Percentage Error (SMAPE) were reduced by 12.3%, 13.5%, and 6.1%, respectively.

Conclusions:

• DiffMixer effectively analyzes and predicts different frequencies in non-stationary time series.
• The proposed method outperforms state-of-the-art techniques in forecasting accuracy.
• This approach offers a robust solution for real-world time series prediction challenges.