Continual Learning for MTS Forecasting brings a novel approach to address the challenges of multivariate time series (MTS) forecasting under a continual learning paradigm.

SKI-CL framework steers forecasting models to adapt to different regimes based on a graph structure learning backbone, which incorporates a consistency regularization with structural knowledge.

• Unveils a representation-matching memory replay mechanism maximizing temporal coverage for MTS data.
• Offers impressive results in terms of forecasting accuracy and the ability to maintain performance over multiple regimes.
• Provides empirical studies demonstrating the benefit of SKI-CL over state-of-the-art strategies.

The success of this framework emphasizes the importance of integrating various domain-specific knowledge and memory mechanisms in continual learning systems, offering a strategic edge in MTS forecasting.