Hybrid evolutionary-gradient training improves long-term time series forecasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Time Series Analysis

Background:

• Long-term time series forecasting faces challenges like nonstationarity, noisy gradients, and distribution shifts.
• Existing methods often struggle with robust learning under these conditions, leading to delayed adaptation and reduced accuracy.

Purpose of the Study:

• To introduce an architecture-agnostic training framework, Evolutionary-Guided Module Fusion with Gradient Refinement (EGMF-GR).
• To enhance the robustness and stability of long-term time series forecasting models.

Main Methods:

• EGMF-GR integrates population-based global exploration with gradient-based local refinement.
• It monitors module alignment and discrepancies between individuals, using hybrid thresholds for module state fusion.
• Fusion occurs at the module state level, merging parameters and synchronizing buffers to ensure stability.

Main Results:

• EGMF-GR demonstrated improved forecasting accuracy across eight public benchmarks.
• The framework significantly enhanced training stability, particularly under challenging conditions.
• Optimization stability was improved by reducing state inconsistency after module merging.

Conclusions:

• EGMF-GR offers a novel approach to improving long-term time series forecasting.
• The framework effectively addresses nonstationarity and distribution shifts without requiring new forecasting architectures.
• EGMF-GR provides a stable and accurate training strategy within a controlled optimization budget.