Researchers reveal how AI explainability unlocks adaptive time series forecasting

A new paper from French research teams (RFAI, BDTLN, LIFAT) confronts the dual challenges of non-stationarity and black-box behavior in time series forecasting. The authors deployed three modern neural architectures—PatchMixer, PatchTST, and DLinear—augmented with attention-based sampling mechanisms for continual learning via Experience Replay. To peer inside these models, they applied explainability techniques: attention rollout and Grad-CAM (gradient-based attribution). Experiments were conducted on real-world piezometric (groundwater level) time series that exhibit heterogeneous patterns and regime shifts.

Beyond raw predictive accuracy, the study reveals how attribution patterns shift as the model adapts. These evolving attributions help identify when and why the model's focus changes, informing better data selection and adaptation strategies. The work, presented at EDBT/ICDT 2026, underscores that explainability is not just a post-hoc curiosity but a practical lever for designing robust continual learning systems in domains like environmental monitoring.