An Explainable AI Framework for AC Optimal Power Flow

Abstract

Despite their success and impressive performance, deep learning models are still considered opaque models. Experts have minimal trust in adopting them when operating critical systems, such as power systems. Integrating renewable energy sources increases grid uncertainty, making it crucial to frequently solve the Alternating Current Optimal Power Flow (AC-OPF) problem to provide cost-effective and reliable grid operations. This paper proposes an interpretable deep learning framework to address the AC-OPF problem, offering explainability without sacrificing accuracy. Feature attribution techniques, including Expected Gradients and Shapley Additive Explanations (SHAP), are employed to identify the most influential input features on the outputs, both individually and collectively. Interpretability is essential for building trust for system operators and plays a key role in debugging, validation, and fine-tuning. The proposed approach is evaluated on four test grids, demonstrating high accuracy and interpretability. The results confirm that the model aligns with the physical principles of power systems, providing trustworthy and explainable decision-making to power system operators.