Explainability Challenges in Continual Learning for Time Series Forecasting

Deep learning models show promise for time series forecasting, but their real-world deployment, especially in dynamic environments like environmental monitoring, is hindered by non-stationary data and a lack of explainability. This study investigates how explainability can illuminate the behavior of continual learning systems in adaptive time series forecasting, specifically those employing experience replay strategies. The researchers applied attention rollout and gradient-based attribution methods, such as Grad-CAM, to analyze neural forecasting architectures like PatchMixer, PatchTST, and DLinear. Experiments on real-world piezometric time series, which exhibit diverse patterns and regime shifts, demonstrated that analyzing both predictive behavior and sampling strategies provides crucial insights into the dynamics of continual learning. The findings highlight the complexities and opportunities of using explainability to understand how attribution patterns change over time, which can inform better data selection and adaptation strategies in constantly evolving forecasting scenarios.