Learn Dynamic Hosting Capacity Based on Voltage Sensitivity Analysis

The extensive use of distributed energy sources (DERs) presents the substantial design, planning, and operational issues for distribution systems, thus prompting the broad adaption of methodologies for photovoltaics (PV) hosting capacity analysis (HCA). Traditional HCA methods require running power flow analysis iteratively, typically in the time-series scenario, to consider the dynamic pattern. However, the time-consuming HCA techniques fail to offer online prediction in large distribution networks because of the computational burden. To tackle the computation challenge, we first provide a deep learning-based problem formulation for HCA, which performs offline training and calculates hosting capacity in real time. The applicable learning model, long short-term memory (LSTM), uses historical time-series data to identify the underlying periodic patterns in distribution systems. However, the accuracy of HC estimation is low in the LSTM without considering system spatial information correlated with HC. To capture such spatial correlation from system measurements, we design dual forget gates in the LSTM and propose a novel Spatial-Temporal LSTM. Moreover, as voltage violations are observed to be one of the most critical constraints of HCA, we construct a voltage sensitivity gate to increase the weight on voltage variation and reduce the mismatch in HC determination. The simulation results on different feeders, such as IEEE 123-bus and utility feeders, validate our designs.