Physics-Guided Neural Network for Load Margin Assessment of Power Systems

Abstract

The power system load margin is an important index used in power system operation centers. Traditional load margin calculation involves solving a set of differential-algebraic equations where not all information is always available. Data-driven methods such as Neural Networks (NNs) have been proposed but present poor performance and lack of generalization, especially in cases other than the operating history of power systems. In this article, we propose a Physics-Guided Neural Network (PGNN) to calculate the load margin of power systems. The PGNN is regularized by an auxiliary process that aims to reconstruct the power flow equations at the stability threshold that defines the load margin. In the PGNN training stage, there are two error functions that must be minimized, one related to the empirical knowledge of the system and the other related to the physical knowledge of the system associated with the guarantee of power flow. Case studies are presented for two test power systems, IEEE 68-bus system and Brazilian interconnected power system, considering a set of operating conditions, signal noise and outliers. The results show the superior performance of PGNN over traditional NNs.