Integrated Data-Driven Framework for Rooftop PV Impact Assessment in Distribution Networks Using Net-Load Forecasting and Hosting Capacity Analysis

The increasing use of rooftop photovoltaic (PV) systems in distribution networks can lead to operational challenges, including feeder overloading, reverse power flow, fluctuating net-load characteristics, and future hosting-capacity constraints. This research presents an integrated data-driven approach for the holistic evaluation of the impacts of rooftop PV on the Sohar Grid Station (GSS) distribution network in Oman. The proposed framework integrates spatial feeder-loading analysis, PV adoption assessment, net-load and duck-curve evaluation, reverse-power-flow detection, PV performance analysis under dust conditions, machine learning-based net-load forecasting, hosting-capacity screening, and SHAP-based explainability. The framework is based on operational feeder data, Sahim PV installation and export data, meteorological factors, irradiance, and dust-related metrics. The findings demonstrate that the present network is already under operational stress, with 10 out of 35 feeders exceeding the 100% loading reference limit, including two feeders that reached loading levels of 192.8% and 183.3%, respectively. The present PV penetration of the Sahim system causes a modest decrease in net daytime demand and localised reverse-power-flow effects on feeders with higher PV-to-load ratios. The hosting-capacity study identified a PV accommodation potential of 27.11 MWp DC, compared with an existing installed capacity of 2.037 MWp DC, indicating substantial screening-level potential for additional PV deployment, although the available capacity varies across feeders and should not be interpreted as a definitive interconnection limit. Among the evaluated forecasting methods, the combined convolutional neural network and long short-term memory (CNN-LSTM) model achieved the lowest root mean square error (0.063 MW) and the highest coefficient of determination (0.984). In addition, SHAP analysis of the Random Forest model showed that recent and weekly historical net-load values were the dominant predictors. The proposed framework provides a useful decision-support tool for reliable rooftop PV integration and distribution network planning.