Advanced GPT Inverter Physical Demonstration (AGIPDem)

Transitioning to a sustainable energy society and mitigating climate change effects through accelerated energy sector decarbonization is a recognised top priority of the European Union (EU) under the recent European Green Deal (EGD). Artificial intelligence (AI) is a key enabling technology towards an optimally managed renewable-powered energy sector. The aim of the proposed project was to examine the capabilities of next-generation AI-driven grid-supportive tools that facilitate the dynamic and cost-effective management of microgrids at high shares of solar photovoltaics (PV). Specifically, the project focused on the development and accelerated validation of a data-driven voltage state estimator and a grid condition prognostic platform (dig ital twin) that includes a net-load forecasting model. The activities further entailed the validation of the proposed tools' while coupled to grid supportive controls demonstrated through real-time simulation environments ad control hardware-in-the-loop simulations. To this end, the pro posed project is timely pertinent by offering a real-time simulation framework for experimentally validating smart grid analytical tools and demonstrating AI-driven applications (voltage state estimation and net-load forecasting).

The project was performed cooperatively, in two phases:

• A first physical stay of staff members of University of Cyprus was held at AIT Austrian Institute of Technology at the SmartEST Laboratory in Vienna, Austria.
• A second physical stay was performed 2 weeks later by AIT staff members at University of Cyprus in Nicosia at the Low Voltage Experimental Microgrid Lab, Cyprus to imple ment the developed methods on site.

In the scope of this work, the test system considered for the real-time simulations and the actual site demonstration is the low-voltage (LV) Experimental Nanogrid of UCY-FOSS and the Smart Energy Campus microgrid of UCY. The nanogrid pilot is a flexible and scalable renewable to grid integration infrastructure that includes PV systems, smart inverters, battery storage, smart loads/plugs, smart meters, IoT communication devices and a central energy management system. Along this context, the Smart Energy Campus is a commercial-scale University campus microgrid that comprises of 15 smart buildings and distributed PV systems of capacity 400 kW that are fully monitored and equipped with smart meters as part of the implemented advanced metering infrastructure (AMI) for the acquisition of high-resolution data.

Three test cases were designed for the project:

• Test Case 1 [AIT]: Voltage state estimation tool development and validation for utility scale microgrids (developed and provided by the applicants) at real-time environments.
• Test Case 2 [AIT]: Short-term net-load forecasting voltage regulation tool development and validation for utility-scale microgrids (developed and provided by the applicants) at real-time environments.
• Test Case 3 [UCY-FOSS]: Grid-condition prognostic digital twin (developed and pro vided by the applicants) verification for utility-scale microgrids.

As a result of the conducted tests, the performance of the developed voltage state estimator was verified to achieve high accuracies <1% error, when supplied with high-resolution data (high variability solar and demand data) in a software-in-the-loop approach. In addition, the capability of the voltage state estimator to estimate and follow voltage deviations when imput ing random grid faults was validated using customised SCADA HIL dashboards. An additional optimisaiton step was carried out in order to further imporve the accuracy of the state estimator. For this purpose, the model was trained using synthetic training data (obtained by performing a power flow analysis of the developed UCY microgrid PowerFactory model) and historic measurements. The evaluation results showed that the devised model, leveraging artificial neural networks (ANNs), exhibited high accuracies and was capable to follow in most cases the actual voltage profiles even at low bus-bur utilisation fractions.

The tests further verified the performance accuracy of the optimally constructed net-load fore casting model that yielded forecasting errors of approximately 4%. Moreover, the provision of grid control functionalities through the real-time simulation model (driven by the implemented tools) was emulated using the AIT SGC with SunSpec inverter protocol support.

Finally, the performed research is expected to enhance the predictive observability and prog nostic control of smart grids/microgrids leading to increased grid flexibility for integrating higher shares of PV at the distribution network. Actual-life demonstrations of such intelligent tools are therefore invaluable for grid operators that aim to optimally orchestrate complex distribution system operations at high-RES shares.