Explainable Load Forecasting with Covariate-Informed Time Series Foundation Models
| Authors | Matthias Hertel et al. |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2604.28149 |
| Download | |
| Categories | cs.LG |
Abstract
Time Series Foundation Models (TSFMs) have recently emerged as general-purpose forecasting models and show considerable potential for applications in energy systems. However, applications in critical infrastructure like power grids require transparency to ensure trust and reliability and cannot rely on pure black-box models. To enhance the transparency of TSFMs, we propose an efficient algorithm for computing Shapley Additive Explanations (SHAP) tailored to these models. The proposed approach leverages the flexibility of TSFMs with respect to input context length and provided covariates. This property enables efficient temporal and covariate masking (selectively withholding inputs), allowing for a scalable explanation of model predictions using SHAP. We evaluate two TSFMs - Chronos-2 and TabPFN-TS - on a day-ahead load forecasting task for a transmission system operator (TSO). In a zero-shot setting, both models achieve predictive performance competitive with a Transformer model trained specifically on multiple years of TSO data. The explanations obtained through our proposed approach align with established domain knowledge, particularly as the TSFMs appropriately use weather and calendar information for load prediction. Overall, we demonstrate that TSFMs can serve as transparent and reliable tools for operational energy forecasting.
Engineering Breakdown
Plain English
This paper addresses a critical gap in deploying Time Series Foundation Models (TSFMs) to critical infrastructure like power grids: these models are black boxes that lack transparency, making them unsuitable for systems requiring trust and reliability. The authors propose an efficient algorithm to compute SHAP (Shapley Additive Explanations) values specifically for TSFMs, exploiting the models' ability to handle variable input context lengths and selective masking of covariates. They evaluate their approach on two foundation models (Chronos-2 and another unnamed model) to demonstrate scalable, interpretable predictions. This work makes TSFMs practically viable for high-stakes applications where regulators and operators need to understand why the model made specific forecasting decisions.
Core Technical Contribution
The core novelty is an efficient SHAP computation algorithm tailored to the architectural properties of TSFMs rather than treating them as generic black boxes. Prior SHAP implementations require extensive model re-runs with different input subsets, which becomes prohibitively expensive for large foundation models. The authors leverage two key SHAP-friendly properties of TSFMs: (1) flexible input context length, allowing variable-length historical windows, and (2) native support for selective covariate masking without retraining. This enables computing feature importance and temporal importance scores with significantly fewer forward passes than standard SHAP, making explanation tractable at production scale.
How It Works
The algorithm operates by systematically masking temporal windows and covariate subsets in the TSFW input, then measuring how prediction accuracy degrades when each element is removed. Input to the TSFW is a historical time series window plus optional external covariates (weather, pricing, etc.); the foundation model processes this context and outputs a forecast. Instead of retraining or fine-tuning, the approach leverages the fact that TSFMs can accept shorter context windows or zeroed-out covariates without architectural changes. For each coalition of features/time-steps to evaluate, the model processes a masked input and compares the resulting prediction to the baseline; aggregating these comparisons across all possible coalitions yields Shapley values that attribute prediction confidence to specific inputs. The efficiency gain comes from avoiding the combinatorial explosion of model variants—a standard SHAP baseline might require 2^n forward passes for n features, but the flexible masking reduces redundant computation by reusing the same model weights.
Production Impact
For engineers deploying TSFMs in energy grids or other regulated critical infrastructure, this brings interpretability from a research curiosity to a production requirement. Regulatory compliance in power systems often mandates that operators understand model decisions—FERC, NERC, or national grid operators will demand explanations for forecast-driven dispatch decisions. Without this work, teams would face a choice: use opaque TSFMs (higher accuracy, black box) or stick with interpretable statistical models (lower accuracy, explainable). Now teams can have both: deploy a TSFW with built-in explainability at reasonable computational cost. The trade-off is moderate: computing SHAP explanations for a single forecast requires multiple forward passes (exact cost depends on feature/temporal granularity), adding latency of perhaps 5-10x a single inference, but this is amortized over batch processing and can run offline. Integration is straightforward—wrap the TSFW encoder with a masking layer, then compute importance scores post-hoc without modifying the model itself.
Limitations and When Not to Use This
The paper does not address how to handle non-linear or interaction-heavy feature relationships—SHAP assumes additive feature contributions, which may not hold when covariates interact strongly (e.g., wind speed and temperature jointly affect solar generation). The evaluation is limited to two foundation models, leaving open whether the efficiency gains generalize across architectures or smaller domain-specific TSFMs. The approach assumes masking with zeros is a reasonable counterfactual, but for time series data this may be unrealistic—real grid conditions never have missing covariates, so SHAP explanations may conflate absence with genuine feature importance. Finally, the paper does not benchmark against simpler post-hoc explanation baselines (e.g., attention weights, gradient-based attribution) which might be faster or comparably interpretable in practice; true computational savings are unknown without direct comparison.
Research Context
This work sits at the intersection of three active research areas: (1) foundation models for time series (Chronos, Lag-Llama, and similar recent models aim for general-purpose forecasting), (2) interpretability methods for deep learning (SHAP is the standard, but expensive for large models), and (3) AI for critical infrastructure (energy, power systems, supply chains demand transparency). The paper builds directly on SHAP's additive framework but optimizes it for the specific architectural properties of modern TSFMs that prior interpretability work did not exploit. It extends the emerging practice of adapting foundation models to domain-specific tasks by ensuring that domain adoption (e.g., in power grids) includes explanation capabilities from the start rather than bolting them on later. The work opens future directions in temporal attribution (understanding which past time steps matter most for a forecast) and covariate importance ranking, which could inform sensor placement and data collection strategies in real grids.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::