Preventive-corrective contingency control of distribution power grids using convolutional neural networks

Abstract

To achieve the goals on greenhouse gas emissions, the energy supply and demand is in transition. Distribution power grids therefore are increasingly reaching their capacity limits due to electrification and the vast increase of distributed energy resource (DER) connection requests with large peak power output. Increasing physical grid capacity is a costly operation and take lots of time to realise. Grid operators are therefore allowed to connect additional energy resources to the power grid at the cost of N-1 reserve capacity. Having N-1 reserve capacity means that any grid component can go out of service, without causing overloading of another grid component. When this security principle is abandoned, coupled preventive and corrective control measures might be necessary to preserve security of energy supply.
In this thesis, commissioned by the Dutch distribution grid operator Stedin Netbeheer, a preventive-corrective contingency control method based on differential evolution (DE) is designed to increase the maximum admissible DER generation on a distribution grid. For the contingency analysis process in this contingency control method, the full AC power flow method is compared to the method based on line outage distribution factors. The resulting DE-based preventive-corrective control method using the full AC power flow contingency analysis method allows for significant extra DER generation capacity to be connected to the study case distribution power grid, without requiring expensive grid expansions. The case study of this thesis work is the Stedin Middelharnis distribution power grid. Due to the computational complexity of the preventive-corrective contingency control method and the sparse connectivity of power system data, convolutional neural networks are deployed to reduce the computational time. In the first approach, the convolutional neural network is used to perform contingency analysis. Due to this neural network approach, the time to compute the preventive-corrective control actions is reduced by 40%. However, the accuracy of the control is significantly reduced due to the inaccuracy of the used contingency analysis method. In the second approach, a neural network is trained to determine the coupled preventive control actions. The performance of a neural network with and without convolutional neural networks is compared in this approach. Results show that the convolutional neural network outperforms the neural network without convolutional layers. This second convolutional neural network approach is satisfactory accurate and the computational efficiency of this control method is greatly increased compared to the control method based on DE, making realtime preventive-corrective control possible.