A Novel Contingency Analysis-Driven Deep Learning Framework for Evaluation of a Primary Power System Security Metric

Continuous operation of power systems is critical to modern society, even during unexpected or emergency situations. One method to evaluate the security of power systems is contingency analysis, a computationally-intensive process which considers how a set of possible scenarios would affect the electrical grid. This paper presents a novel security evaluation method which enables a topology- and parameter-free estimation of a system-level security metric, the system aggregate megawatt contingency overload (SysAMWCO), by leveraging a novel deep learning-based approach. Using the results of conventional contingency analysis for a simulated power system, a deep neural network (DNN) is trained to predict the SysAMWCO value using only bus loads as input. Experimentation across different neural network parameters reveals a complex relationship between these parameters and the accuracy of the predicted SysAMWCO value. While each DNN is trained on data generated using conventional contingency analysis methods, fully-trained models achieve R-squared scores in excess of 0.98 without needing electrical grid topology or power device parameter information as input. Given these results, this framework shows promise for potential real-world applications.