Modeling Static Grid Equivalent with Artificial Neural Networks Including Optimally Designed Local Reactive Power Control

During the energy system transition, the rapid increase in distributed energy resources (DERs) can cause grid congestion problems in power grids, such as voltage problems and overloads. It has become increasingly important to intelligently control the provision of reactive power from DERs to power grids, which not only stabilizes voltage, reduces loading and grid losses, but also provides reactive power flexibility to different points in power grids. As power grids are interconnected and operated by different grid operators, using an equivalent grid model in place of a complete detailed model is the standard solution for data confidentiality and computational efficiency. The equivalent model should represent the behavior of the original grid as accurately as possible. However, current state-of-the-art equivalent grids are based on a static grid state, meaning that variable grid states, such as those caused by strong power fluctuations, controllers, and switching state changes, are often not represented by the equivalent model. This can lead to significant errors in power system analysis.
This thesis proposes a time series optimization-based method for calculating optimal characteristic curves for each DER, enabling intelligent local reactive power control. The open-source tools pandapower and PowerModels are functionally interconnected to optimize reactive power provisions of DERs, and linear decision tree regression is used to identify the optimal characteristic curve from the optimization results. To consider variable grid states resulting from the factors mentioned above, a static grid equivalent method based on artificial neural networks (ANNs) is developed. The use of ANNs with feedforward and recurrent architectures significantly improves the accuracy of the grid equivalent, compared to the state-of-the-art methods. Furthermore, to ensure data confidentiality, an unsupervised ANN - an Autoencoder - is implemented. It obfuscates the original grid data for exchange while preserving its features for training with sufficient accuracy.
The proposed methods are analysed and evaluated based on extensive simulations with different grids and representative scenarios. The individually optimized characteristic curves are found to be more effective than conventional settings in supporting voltage stability, reducing grid losses, and providing reactive power flexibility. They can serve as a backup for failures in central optimizations or as a bridging solution for grid operators without modern grid monitoring systems. By incorporating the characteristic curve parameters and the switching status as training data and using the recurrent architectures, the ANN-based equivalent accuracy is further improved on the existing improvements with the feedforward architecture. Compared to the state-of-the-art REI equivalent, the equivalent deviations are reduced by approximately 90%. The entire exchange and training process is based on the obfuscated grid data, ensuring data confidentiality.