Using machine learning to advance computational efficiency in stochastic hosting capacity evaluations

Assessment of distribution networks' hosting capacity (HC) is critical for optimal integration and utilization of distributed energy resources (DERs). DERs have various uncertainties which must be captured for accurate HC assessment. Stochastic hosting capacity (SHC) methods allow comprehensive and accurate HC analysis given their ability to simulate a wide range of input scenarios. However, due to the need of load flow assessment for each input scenario, SHC methods have extensive computational requirements. Various computational efficiency improvements in SHC analysis have been proposed but the use of machine learning (ML) in such applications is still limited. This paper presents an efficient ML-driven SHC (SHC-ML) method based on linear regression. The proposed method utilizes a limited scenario set to train an ML model that accurately and efficiently predicts the voltage impact of a comprehensive, larger input scenario set. This approach significantly reduces the number of load flow computations in the SHC assessment. The proposed method is demonstrated on three low-voltage distribution feeders with rooftop photovoltaic systems. The results show that a significantly reduced subset as low as 0.003
^%
of the standard SHC scenario set can sufficiently train an SHC-ML model to accurately predict voltage impacts given DER penetration scenarios. The reduction in the conducted load flow assessments improves the computational efficiency of the SHC analysis by a factor of up to 17, with high accuracy (>99
^%
). The results have significant implications for the optimal integration and utilization of DERs.