Data-driven AC Optimal Power Flow with Physics-informed Learning and Calibrations

Junfei Wang,Pirathayini Srikantha

2024-08-29

Abstract:The modern power grid is witnessing a shift in operations from traditional control methods to more advanced operational mechanisms. Due to the nonconvex nature of the Alternating Current Optimal Power Flow (ACOPF) problem and the need for operations with better granularity in the modern smart grid, system operators require a more efficient and reliable ACOPF solver. While data-driven ACOPF methods excel in directly inferring the optimal solution based on power grid demand, achieving both feasibility and optimality remains a challenge due to the NP-hardness of the problem. In this paper, we propose a physics-informed machine learning model and a feasibility calibration algorithm to produce solutions for the ACOPF problem. Notably, the machine learning model produces solutions with a 0.5\% and 1.4\% optimality gap for IEEE bus 14 and 118 grids, respectively. The feasibility correction algorithm converges for all test scenarios on bus 14 and achieves a 92.2% convergence rate on bus 118.

Systems and Control

What problem does this paper attempt to address?

The paper aims to address the Alternating Current Optimal Power Flow (ACOPF) problem in modern power systems. Specifically, the authors propose a physics-informed deep learning framework to generate optimal solutions for the ACOPF problem that satisfy all physical constraints. Traditional ACOPF problems are non-convex and have been proven to be NP-hard, making efficient solving challenging. Additionally, with the increase in renewable energy, unpredictable loads, and the proliferation of electric vehicles, modern power grids require faster and more reliable ACOPF solvers.

The paper presents the following two main contributions:
1. **Physics-Informed Deep Learning Model**: Utilizing deep neural networks (DNN) to predict the optimal voltage magnitudes and phase angles for a given power demand, the model is trained by minimizing prediction errors and power injection reconstruction loss based on physical information.
2. **Feasibility Correction Algorithm**: An iterative correction algorithm is employed, using the Gauss-Seidel update method to make local adjustments at load nodes and directly adjusting power injections at generator nodes to eliminate any feasibility-related errors.

Experimental results show that on the IEEE 14-bus and 118-bus systems, the proposed method achieves near-optimal solutions and maintains a high convergence rate during correction, at 100% and 92.2% respectively. These results outperform existing data-driven ACOPF algorithms. Future work will explore robust machine learning methods under input uncertainty, unsupervised or semi-supervised learning, and conditions for theoretically guaranteed convergence.