Data-driven photovoltaic hosting capacity calculation method for unbalanced low voltage distribution networks

• Explicit neural networks estimate LVDN voltage from smart-meter data without needing physical parameters. • Adaptive partitioning via networks sensitivity enables efficient node division under uncertain LVDN topology. • PSO-optimized PV hosting models with power quality constraints ensure efficient computation and safe integration. Distributed photovoltaic (PV) hosting capacity is commonly used to assess the PV capacity allowed to be accommodated under the constraints of safe and stable operation. However, the existing methods rely on accurate power flow, which is difficult to be applied to low voltage distribution networks (LVDNs) with unclear topology, inaccurate line parameters, and mixed phase sequence access. Therefore, this paper proposes a data-driven calculation method for distributed PV hosting capacity in LVDNs. Firstly, an explicit neural network is applied to fit the mapping relationship between the power of each node and the voltage of nodes in LVDNs, and construct explicit equations reflecting the connection between the power of the nodes and the voltage of the nodes according to the inference logic of the network. Secondly, a clustering performance index is constructed by taking into account the sensitivity of the phase voltages to the node power in order to realize the adaptive partition of the LVDNs for multi-phase user access. Finally, based on the clustering performance index, a hosting capacity calculation model is constructed for the LVDNs that takes voltage deviation and three-phase unbalance into account. The IEEE-123 node model and an actual distribution network are verified and the results show that the proposed hosting capacity method is not only applicable to the actual LVDNs scenario but also has excellent robustness, which can approach the theoretical hosting capacity of the LVDNs under some observable conditions.