Implementation of Machine Learning-based DER Local Control Schemes on Measurement Devices for Counteracting Communication Failures

One of the significant challenges linked with the massive integration of distributed energy resources (DER) in the active distribution grids is the uncertainty it brings along. The grid operation becomes more arduous to avoid voltage or thermal violations. While the Optimal Power Flow (OPF) algorithm is vastly discussed in the literature, little attention has been given to the robustness of such centralised implementation, such as the provision of redundant control solutions during a communication failure. This paper aims to implement a machine learning-based algorithm at each Intelligent Electronic Device (IED) that mimics the centralised OPF used during communication failures using IEC 61850 data models. Under normal circumstances, the IEDs communicate for centralised OPF. In addition, the system is trained offline for all operational conditions and the individual look-up tables linking the actual voltages to the DER setpoints are sent to the respective controllers. The regression models allow for the local reconstruction of the DER setpoints, emulating the overall OPF, in case of a communication failure. In addition to the regression control, the paper also explains an offline learning approach for periodic re-training of the regression models. The implementation is experimentally verified using a Hardware-in-the-loop test setup. The tests showed promising results compared to conventional control strategies during communication failures. When properly trained and coordinated, such an intuitive local control approach for each DER could be very beneficial for the bulk power system. This machine learning-based approach could also replace the existing Q(V) control strategies, to better support the bulk power system.