SHAP-Guided Hybrid Deep Learning Enhances Grid Load Forecasting Under Extreme Weather
A recent preprint on arXiv (2604.23500) presents a new ensemble framework that uses insights from physics to improve short-term electricity load predictions and strengthen the reliability of the U.S. grid during extreme weather. This innovative model combines a Convolutional Neural Network (CNN) to capture local features with a Transformer to handle long-range dependencies, brought together through a carefully weighted ensemble optimized through validation. To ensure effective training, it utilizes a physics-informed loss based on the temperature-demand relationship specific to the Electric Reliability Council of Texas (ERCOT). Additionally, it employs SHapley Additive exPlanations (SHAP) for interpretability, leveraging data from ERCOT and Automated Surface Observing System (ASOS) stations from 2018 to 2025.
Key facts
- Proposes interpretable physics-informed ensemble for short-term load forecasting
- Integrates CNN and Transformer branches with validation-optimized weighted ensemble
- Physics-informed loss uses ERCOT's piecewise parabolic temperature-demand relationship
- SHAP with DeepExplainer provides post-hoc interpretability
- Validated on eight years of ERCOT hourly load data (2018–2025)
- Uses ASOS records from three stations
- Aims to improve U.S. grid reliability during extreme weather
- Addresses opacity of deep learning models limiting operator trust
Entities
Institutions
- Electric Reliability Council of Texas (ERCOT)
- Automated Surface Observing System (ASOS)
Locations
- United States
- Texas