PHYSICS-INFORMED GENERATIVE ARTIFICIAL INTELLIGENCE FOR MODELING, PREDICTION, AND OPTIMIZATION OF SMART ELECTRICAL SYSTEMS

Abstract

Purely data-driven artificial intelligence models for smart electrical systems suffer from the need for large amounts of labeled data, the lack of guaranties on physically meaningful outputs, and the tendency to produce predictions and scenarios that violate the laws governing power networks . As grids are incorporating larger shares of distributed and uncertain resources, these shortcomings are becoming increasingly consequential. In this work, we propose a physics-informed generative artificial intelligence framework that unifies the modeling, prediction and optimization of smart electrical systems in a single architecture. The framework integrates a conditional denoising diffusion model that captures the joint uncertainty of loads, renewables, and network states, with a differentiable physics layer that embeds the underlying electrical laws, i.e., AC power-flow equations, voltage and branch limits, and system power balance, directly into the learning objective. The training is based on a composite loss function that combines a data loss, a physics constraint violation term and a regularization term with tunable weighting factors. This allows the model to generate scenarios that are both statistically faithful and physically consistent. The generated scenario ensembles serve as a probabilistic prediction engine and a scenario based stochastic optimization module. The physics layer verifies and projects the outputs on the feasible manifold during inference. Key findings show that the physics-informed formulation offers a beneficial inductive bias that improves generalization in data-scarce regimes, substantially reduces constraint violations, and limits physically implausible drift in long-horizon forecasts. The generative module produces full joint scenario ensembles and effective data augmentation, and the optimization module reaches near-optimal decisions at a significantly lower computational cost than traditional iterative solvers. Ablation studies show that these physics constraints and the generative mechanism are complementary to each other in providing gains. The framework is of practical value for day-ahead and intraday probabilistic forecasting, renewable scenario generation, synthetic data production under privacy constraints, and dispatch and market operations. Future work will include foundation model architectures, explicit feasibility guaranties, multi-energy extensions and real-time digital twin integration.

Keywords: Physics-informed AI; generative AI; smart electrical systems; diffusion models; probabilistic forecasting; scenario generation; power system optimization; constraint satisfaction; deep learning.