Contingency Analysis of Power Systems with Artificial Neural Networks

Abstract

A fast assessment of the single contingency policy for power systems is crucial in power system planning and live operation. Power system planning methods based on thousands of power flow calculations, such as time series based grid planning strategies, rely on a fast evaluation of loadings in case of simulated outages. Standard approximation methods, such as the line outage distribution factor (LODF) matrix, have limited accuracy and can only approximate real power flows. To increase accuracy and to predict other power system parameters, we perform contingency analysis with artificial neural networks. Deep feedforward network architectures are trained with 20% of AC power flow results from time series simulation of one year. The remaining line loadings and bus voltages are then predicted. Detailed analyses are conducted on a real German 110 kV sub-transmission grid located in Karlsruhe. The method is additionally tested on the IEEE57 bus system and the CIGRE15 bus medium voltage grid. For each test grid prediction errors are extremely low (0.5%) in comparison to the LODF method (18.6%). Prediction times are significantly less compared to AC power flow calculations (10s vs. 1861s).