Due to climate targets of the German government, the share of renewable energy in the power grid will be increased and the number of grid participants connected to the low voltage level of the power grid will rise. This leads to new requirements in voltage control, especially in low voltage distribution grids. In order to achieve a stable power grid in future, further development of control strategies is necessary. In this paper, a load recognition concept, which was tested on simulative data in previous work, is further developed to reduce simulation effort. Additionally, the concept is adapted for real hardware influences and active grid participants complicating the recognition task. Thus, the main contribution of this study is the successful application of the methodology within a hardware-based test grid containing a charging electric vehicle. Using a convolutional neural network in a time series classification setting, the recognition rates in this use-case exceeded 99 % while benefiting from an asymmetric charging behavior. Due to these promising results, future voltage control strategies could be supported based on gained information through integration of the presented concept.

Due to the increasing penetration of renewable energies in lower voltage level, there is a need to develop new control strategies to stabilize the grid voltage. For this, an approach using deep learning to recognize electric loads in voltage profiles is presented. This is based on the idea to classify loads in the local grid environment of an inverter’s grid connection point to provide information for adaptive control strategies. The proposed concept uses power profiles to systematically generate training data. During hyper-parameter optimizations, multi-layer perceptron (MLP) and convolutional neural networks (CNN) are trained, validated, and evaluated to determine the best task configurations. The approach is demonstrated on the example recognition of two electric vehicles. Finally, the influence of the distance in a test grid from the transformer and the active load to the measurement point, respectively, onto the recognition accuracy is investigated. A larger distance between the inverter and the transformer improved the recognition, while a larger distance between the inverter and active loads decreased the accuracy. The developed concept shows promising results in the simulation environment for adaptive voltage control.

Due to the increasing penetration of the power grid with renewable, distributed energy resources, new strategies for voltage stabilization in low voltage distribution grids must be developed. One approach to autonomous voltage control is to apply reinforcement learning (RL) for reactive power injection by converters. In this work, to implement a secure test environment including real hardware influences for such intelligent algorithms, a power hardware-in-the-loop (PHIL) approach is used to combine a virtually simulated grid with real hardware devices to emulate as realistic grid states as possible. The PHIL environment is validated through the identification of system limits and analysis of deviations to a software model of the test grid. Finally, an adaptive volt–var control algorithm using RL is implemented to control reactive power injection of a real converter within the test environment. Despite facing more difficult conditions in the hardware than in the software environment, the algorithm is successfully integrated to control the voltage at a grid connection point in a low voltage grid. Thus, the proposed study underlines the potential to use RL in the voltage stabilization of future power grids.

The expansion of renewable energies, the growing number of consumers, e. g., in e-mobility, and their connection to low voltage grids pose new challenges for grid operation, particularly regarding voltage control and energy losses. In this context, comprehensive knowledge about the surrounding grid facilitates the intelligent control of photovoltaic (PV)-battery systems, creating a need to observe grid participants. In that regard, a previously developed load recognition method is successfully applied in a use case to recognize two electric vehicles and a heat pump in a simulative grid environment, yielding an accuracy of around 96%. Considering to couple that method with a subsequent control algorithm, an approach is developed to estimate a virtual voltage signal, which represents the grid voltage without the impact of control actions. It is validated that this enables the effective application of load recognition in control conditions. Finally, it is demonstrated how to integrate the load recognition method into a control setup using the virtual voltage concept. The corresponding control algorithm manages the power flow between a PV-battery system and the grid in conjunction with a static Q(U) algorithm. It leverages the information about active loads with the objective of maintaining a stable and balanced grid voltage and optimizing energy consumption in the example use case. Thereby, the voltage deviation is reduced by around 26% and the grid unbalance by approximately 38%, compared to only using the Q(U) algorithm without a battery. It can be concluded that the load recognition method can successfully gather information from the surroundings of a grid node even in control applications.

Modern electrical grids are evolving into smart, decentralized systems with high integration of distributed energy resources, bidirectional power flows, and real-time control complexity, motivating the need for virtual modeling frameworks. This paper presents a workflow for building and experimentally validating a high-fidelity digital twin of a laboratory-scale reference grid implemented in real hardware. The grid consists of one 100 kVA supply terminal and six 30 kVA load terminals interfaced through a Speedgoat real-time system and monitored via DEWESoft acquisition tools. The digital twin is implemented in MATLAB/Simulink using fixed-step integration and is iteratively refined. Starting from nominal parameters, the model is calibrated by incorporating measured cable impedances and parasitic capacitance derived from the hardware’s behavior. For validation, five representative load profiles are defined and applied in both the simulated and physical grid environments, including static, dynamic, harmonic, and asymmetric conditions. In addition, a coordinated multi-terminal scenario applies varied profiles simultaneously across all six terminals, further confirming the twin’s robustness. The voltage behavior from the hardware is closely reproduced in simulation, with RMS deviations of less than 0.5 V across all scenarios. The twin’s accuracy is demonstrated across multiple loading scenarios, including asymmetrical and harmonics-rich conditions, where it reliably reproduced per-phase voltage dynamics and neutral current behavior observed in hardware. The results confirm that calibrated digital twins can improve development efficiency and reliability in grid research by reducing physical iteration cycles and enabling safe, repeatable evaluation of grid behavior under complex and dynamic conditions.