Skip to content

Explainable Load Forecasting with Covariate-Informed Time Series Foundation Models

Home / Papers / Explainable Load Forecasting with Covariate-Informed Time Series Foundation Models

Explainable Load Forecasting with Covariate-Informed Time Series Foundation Models

Authors: Matthias Hertel, Alexandra Nikoltchovska, Sebastian Pütz, Ralf Mikut, Benjamin Schäfer, Veit Hagenmeyer Date: 2026-04-30 Paper ID: openalex:2604.28149

Summary

This paper addresses the transparency gap in using Time Series Foundation Models (TSFMs) for critical infrastructure by proposing an efficient SHAP-based explanation algorithm. By leveraging the flexible context window and covariate inputs of models like Chronos-2 and TabPFN-TS, the approach enables scalable model interpretability. The framework is evaluated on day-ahead load forecasting, where it demonstrates competitive zero-shot performance and aligns with expert domain knowledge regarding weather and calendar dependencies.

Key Contributions

  • Introduces an efficient SHAP-based explanation algorithm tailored for Time Series Foundation Models (TSFMs) by utilizing their ability to handle varying context lengths and covariates.
  • Demonstrates that Chronos-2 and TabPFN-TS, in a zero-shot setting, achieve performance competitive with domain-specific Transformer models on TSO day-ahead load forecasting.
  • Validates that the proposed explanation framework successfully recovers domain-consistent insights, specifically confirming the importance of weather and calendar features for power load prediction.

Open Questions & Future Work

Key Concepts

  • shap-for-time-series-foundation-models: An algorithm for computing SHAP values for time series foundation models by exploiting their inherent flexibility in input context length and covariate handling.

Archivist Review

I have approved the SHAP-for-TSFM concept and the open question regarding its computational efficiency. These represent a methodological contribution to the intersection of foundation models and interpretability in time series, and a clear, non-trivial research bottleneck. I rejected the generalizability question as it is a typical request for further empirical validation rather than a core research problem.

Approved Concepts

  • SHAP for Time Series Foundation Models: Addresses the critical need for explainability in black-box foundation models applied to high-stakes infrastructure like power grids.

Approved Open Questions

  • Efficient SHAP for TSFMs: Developing more efficient SHAP computation methods is critical for the practical deployment of high-performance TSFMs in real-time, resource-constrained energy infrastructure monitoring.

Rejected Candidates

  • [open_question] Generalizability of TSFM Explainability (generalizability-tsfm-explainability) - low_impact: This is a standard application request for broader evaluation, not a fundamental technical or methodological bottleneck for the vault.

Metadata & Links

url
https://arxiv.org/abs/2604.28149
paper_id
2604.28149
paper_source
openalex
domain
nlp
tags
time-series-question-answering-tsqapatch-based-information-fusion-network
architectures
datasets
concept_slugs
shap-for-time-series-foundation-models
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-03T06:02:28Z