Improving the Accuracy of DC Optimal Power Flow Formulations via Parameter Optimization

TL;DR

This paper introduces a machine learning-inspired parameter tuning algorithm that significantly enhances the accuracy of DC-OPF solutions compared to AC-OPF, achieving up to 90% improvement in certain error metrics.

Contribution

It presents a novel offline parameter optimization method using sensitivity analysis and Quasi-Newton techniques to improve DC-OPF accuracy across various operating conditions.

Findings

Accuracy improvements up to 90% in squared two-norm loss.

Accuracy improvements up to 79% in infinity-norm loss.

Effective offline tuning enhances online DC-OPF solutions.

Abstract

DC Optimal Power Flow (DC-OPF) problems optimize the generators' active power setpoints while satisfying constraints based on the DC power flow linearization. The computational tractability advantages of DC-OPF problems come at the expense of inaccuracies relative to AC Optimal Power Flow (AC-OPF) problems that accurately model the nonlinear steady-state behavior of power grids. This paper proposes an algorithm that significantly improves the accuracy of the generators' active power setpoints from DC-OPF problems with respect to the corresponding AC-OPF problems over a specified range of operating conditions. Using sensitivity information in a machine learning-inspired methodology, this algorithm tunes coefficient and bias parameters in the DC power flow approximation to improve the accuracy of the resulting DC-OPF solutions. Employing the Truncated Newton Conjugate-Gradient (TNC) method, a Quasi-Newton optimization technique, this parameter tuning occurs during an offline training phase, with the resulting parameters then used in online computations. Numerical results underscore the algorithm's efficacy with accuracy improvements in squared two-norm and $\infty$-norm losses of up to $90\%$ and $79\%$, respectively, relative to traditional DC-OPF formulations.