Abstract
Global Forecasting Models (GFMs) provide high accuracy by learning across diverse time series but often function as "black boxes," offering little actionable insight to end-users. We present LIMREF (Local Interpretable Model-agnostic Rule-based Explanations for Forecasting), a framework that generates local, rule-based explanations for complex forecasting models. By combining neighborhood selection, synthetic data generation, and surrogate feature engineering, LIMREF produces "impact rules" and counterfactuals that translate abstract model outputs into human-centric guidance.
Introduction
In energy management, predictive accuracy is insufficient if the underlying reasons for a forecast remain opaque. While GFMs have become the gold standard for accuracy, they lack interpretability. Traditional tools like LIME or SHAP provide feature importance scores that are often too abstract for time-series contexts. A consumer needs to know not just that "Temperature" is important, but how their controllable behaviors—like mean daily consumption—influence their predicted bill. LIMREF addresses this "interpretability-accuracy trade-off" by providing actionable, post-hoc explanations.
The LIMREF Framework
LIMREF operates as a model-agnostic explainer, sitting atop any forecasting architecture. The pipeline involves several key stages:
- Neighborhood Selection: Using Dynamic Time Warping (DTW), LIMREF identifies training instances most similar to the target time series.
- Instance Generation: It employs Moving Block Bootstrapping (MBB) to generate a synthetic neighborhood, allowing the system to probe how the black-box model responds to input perturbations.
- Surrogate Feature Engineering: Raw, high-dimensional data is transformed into interpretable features such as mean consumption, max consumption, and average temperature.
- Rule Mining: The system mines "impact rules" that quantify the effect of specific feature thresholds on the forecast.
Experimental Evaluation
The framework was evaluated using smart meter data from the FUZZ-IEEE 2021 Competition, involving over 3,000 households. The task was to predict monthly electricity demand based on half-hourly readings and exogenous variables like temperature.
Quantitative Performance
LIMREF demonstrated superior accuracy compared to standard interpretable baselines like Linear Regression (LR) and Decision Trees (DT), achieving a Relative Absolute Error (RAE) of 1.259. While LR showed slightly higher fidelity to the black-box model, LIMREF maintained a strong balance between fidelity (0.390 RAE) and human readability.
Qualitative Utility and Rule Generation
The primary strength of LIMREF lies in its ability to generate counterfactual and supporting rules. For example, a rule might state: “If you have 6.01 < average temperature ≤ 6.31 & mean consumption > 14.74... it will increase your consumption by 110.64kWh.” This provides a clear threshold for user action.
Conclusion
LIMREF bridges the gap between high-performance black-box forecasting and actionable decision-making. By focusing on counterfactuals and interpretable surrogate features, it empowers users to understand and influence their predicted outcomes. Future work will explore more granular features and shorter forecasting horizons to further enhance the actionability of the generated rules.
Summary updated