Hosting Capacity Prediction using Spatial-temporal Correlations of Distribution Networks

The widespread deployment of distributed energy resources (DERs), especially photovoltaic (PV), introduces significant challenges in the design, planning, and operation of electric distribution networks. Therefore, PV hosting capacity analysis (HCA) has emerged as a critical planning tool to determine the maximum allowable PV capacity without compromising network performance. However, Traditional HCA methods are constrained by the need for computationally intensive power flow analysis, making them unsuitable for real-time predictions. This study introduces a deep-learning framework for HCA that enables real-time hosting capacity predictions. First, a long short-term memory (LSTM) framework leveraging historical data is used for hosting capacity analysis. However, this approach is not accurate due to not considering topological information. To address this, we have enhanced the LSTM with a dual forget gate mechanism, creating a spatial-temporal LSTM (ST-LSTM) that incorporates both temporal and spatial information. Furthermore, an attention mechanism is integrated to refine the model's focus on specific input data features, improving HC prediction accuracy. Validation conducted on IEEE 123 bus feeders has verified the effectiveness of our method, demonstrating a significant average reduction of 2.35% in prediction errors when utilizing an attention mechanism as opposed to not using it.