Evaluating machine learning models for the fast identification of contingency cases

Abstract

Fast approximations of power flow results are beneficial in power system planning and live operation. In planning, millions of power flow calculations are necessary if multiple years, different control strategies, or contingency policies are to be considered. In live operation, grid operators must assess if grid states comply with contingency requirements in a short time. In this paper, we compare regression and classification methods to either predict multivariable results, for example, bus voltage magnitudes and line loadings, or binary classifications of time steps to identify critical loading situations. We test the methods on three realistic power systems based on time series in 15 and 5 minutes resolution of 1 year. We compare different machine learning models, such as multilayer perceptrons (MLPs), decision trees, k-nearest neighbors, gradient boosting, and evaluate the required training time and prediction times as well as the prediction errors. We additionally determine the amount of training data needed for each method and show results, including the approximation of untrained curtailment of generation. Regarding the compared methods, we identified the MLPs as most suitable for the task. The MLP-based models can predict critical situations with an accuracy of 97% to 98% and a very low number of false negative predictions of 0.0% to 0.64%. K E Y W O R D S contingency analysis, grid planning, machine learning, power flow, time series calculation 1 | INTRODUCTION Power flow (PF) results are the basis for power system planning and are needed in live operation to assess the system state. Quasistatic time series simulations allow evaluating asset loadings, voltage profiles, and contingency situations over a long period, for example, multiple years. This has several advantages in the planning process compared to single "worst-case" analysis including the calculation of grid losses or the integration of demand and generation flexibility. 1,2 However, the computational effort is very high. Millions of PF calculations are necessary if multiple years, different