Efficient Monte Carlo simulation for distributed generation hosting capacity based on statistical distribution analysis

Hosting Capacity (HC) Assessments for Distributed Generation (DG) have been typically carried out using Monte Carlo simulations to cater for different uncertainties, such as, DG sizes and locations, and load consumption patterns. However, their computational cost limits its widespread practical application over real combined unbalanced medium voltage and low voltage (MV-LV) networks. This paper proposes a Statistical Distribution Analysis-based methodology that allows determining a sufficient number of Monte Carlo simulations for HC quantification. This is achieved using a permutation-based analysis for statistical representativeness while mitigating order bias, and employing machine learning (XGBoost) with model explainability (SHAP) to rank the key physical predictors of overvoltage, the most restricting operational constraint in networks with large DG penetration levels. The methodology is demonstrated in a large Chilean MV-LV network with more than ten thousand customers. Results show that a range between 100–200 simulations is sufficient for statistical representativeness under tested error thresholds. Moreover, while network topology metrics consistently dominate, the relevance of photovoltaic (PV) power ratings significantly increases at higher penetration levels. These insights prove restrictive scenarios have identifiable signatures, providing a data-driven foundation for “Physics-Informed” generation of Monte Carlo scenarios, allowing reducing simulation counts and, therefore, computational cost.