The increasing integration of renewable energy sources is fundamentally changing the structure of reactive power provision in distribution networks. While conventional large-scale power plants have traditionally contributed flexibly to voltage control, decentralized generation systems are increasingly taking over this task through Q(U)-characteristics, i. e. voltage-dependent reactive power provisions. In practice, these characteristics are predominantly parameterized statically and without adaptive adjustment to varying grid conditions. The aim of this paper is to develop and validate self-learning, grid-stabilizing and adaptive methods for the optimal parameterization of Q(U)-characteristic curves according to the technical connection rules of distributed renewable energy plants in Germany.

For this purpose, the deterministic regression of optimal power flow results and the heuristic optimization methods Evolutionary Algorithms and Monte Carlo Tree Search are applied to tune each Q(U)-characteristic curve in accourdance to the local grid conditions. In the deterministic approach, the optimal operation points regarding voltage maintenance are determined using the institute’s optimiser eGOpt to perform optimal reactive power flows. Based on these operating points, Euclidean-optimal Q(U)-characteristic curves are derived.

The heuristic methods do not rely on the results of a centralized optimal reactive power flow: For the application of Evolutionary Algorithms, the evolutionary paradigms selection and mutation are utilized to tune the Q(U)-characteristic curve’s parameters. To apply Monte Carlo Tree Search finding an optimal parameterization is translated into a Markov decision process. Further down the decision tree, the differences between neighbouring nodes become smaller and the Q(U)-characteristic curve’s parameters converge to an optimal solution.

The methodology is first validated on a simple distribution network and subsequently applied to a complex power grid comprising numerous generation units and loads to test the scalability of the developed approaches.

The approaches are compared regarding computational effort and voltage maintenance abilities. The results demonstrate that adaptive parameterization improves voltage quality compared to the initial state.