Installed PV capacity detection on LV substations: Comparison of Data-Driven and Model-Based methods

With the proliferation of rooftop Photovoltaic (PV) panels, Distribution System Operators (DSOs) are required to have more detailed awareness of the installed capacity of these assets on their Low Voltage (LV) networks. This paper focuses on different methods of assessing installed PV capacity aggregated at the LV substation level, using data commonly available to DSOs. A framework for a Supervised Data-Driven approach to this problem is proposed, and a comparison is made between these and Model-Based methods, both from the literature and novel. The methods are assessed on their (i) overall performance, (ii) sensitivity to PV penetration, aggregation level (i.e. number of consumers), measurement noise, sampling rate, data seasonality and irradiance errors, and (iii) on a different unseen dataset with real PV generation data. Results show that when highly accurate net load and irradiance data are available, Model-Based methods slightly outperform the Data-Driven ones. However, the Data-Driven methods display higher robustness in scenarios where data quality and availability are not guaranteed, which are realistic operating conditions. Furthermore, the Data-Driven XGBoost outperformed all other methods on unseen, real residential data, where only approximate irradiance data is available. This work highlights the need for careful assessment of the specific conditions present before choosing a method for detecting PV installed capacity. It also validates the possibility of using supervised Machine Learning models for this use case for faster computational times, at the cost of little to no accuracy loss.