CANet: ChronoAdaptive network for enhanced long-term time series forecasting under non-stationarity

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Long-term time series forecasting is crucial for real-world applications.
• Non-stationary data with shifting statistical properties poses significant challenges.
• Existing models often suffer from over-stationarization, limiting performance.

Purpose of the Study:

• To introduce a novel architecture, the ChoronoAdaptive Network (CANet), designed to address non-stationarity in time series forecasting.
• To improve predictive accuracy by preserving essential temporal dependencies and adapting to statistical changes.
• To prevent over-stationarization issues common in current forecasting models.

Main Methods:

• Developed the ChoronoAdaptive Network (CANet) architecture.
• Integrated a Non-stationary Adaptive Normalization module featuring a Style Blending Gate and Adaptive Instance Normalization (AdaIN).
• Employed multi-resolution patching, Fourier analysis-based adaptive thresholding, and a Stacked Kronecker Product Layer.

Main Results:

• CANet demonstrated superior performance over state-of-the-art methods on real-world datasets.
• Achieved a 42% reduction in Mean Squared Error (MSE).
• Achieved a 22% reduction in Mean Absolute Error (MAE).

Conclusions:

• CANet effectively handles non-stationary time series data by preserving and reintegrating statistical characteristics.
• The proposed architecture enhances forecasting accuracy and robustness under dynamic conditions.
• CANet offers a significant advancement in long-term time series forecasting, with source code publicly available.