Dynamic Evaluation of PV Hosting Capacity for Distribution Network based on Deep Reinforcement Learning

The large scale of photovoltaics (PV) integration may exceed the actual hosting capacity (HC) of the distribution network and lead to overvoltage problem. Therefore, it is necessary to carry out a real-time dynamic assessment of photovoltaic hosting capacity (PVHC) to avoid system operation risks. To this end, an online evaluation method of PVHC based on deep reinforcement learning is proposed. A PVHC evaluation model based on the Marko decision process (MDP) theory is first established, which transforms the HC evaluation problem into an sequential dynamic optimal decision problem. Then, an Monte Carlo simulation (MCS) -based soft actor-critic (SAC) algorithm is presented to improve the learning efficiency of the HC evaluation model significantly. The proposed method performs PVHC assessment based on the real-time status of distribution network, and the influence of PV and load uncertainty on the assessment results has been minimized. The numerical analysis of the IEEE 33 bus system shows that the proposed method can achieve the PVHC evaluation results efficiently and accurately.

I. Introduction

The distributed generator is a vital energy choice in the 21st century. It plays an essential role in promoting the process of carbon neutrality. Photovoltaic (PV) has the advantages of clean environmental protection, safety, and reliability. It often exists in the form of distributed PV in the distribution network (DN). However, with the continuous increase of PV permeability, the DN has various problems, such as over-voltage, line overload, and power quality exception^[1]. Therefore, to ensure the coordinated development of DN and PV, it is necessary to evaluate the PV capacity that each bus of the DN can be connected to guide the planning and construction of PV and DN^[2].