Estimation of Aggregated Distribution Grid Flexibilities Using Linear and Nonlinear Machine Learning Algorithms

Inhalt:
Modern power grids are evolving into increasingly complex and decentralized systems due to the integration of distributed energy resources (DERs). The volatility of the DER power infeed introduces fluctuating power flows thereby requiring rapid, real-time solutions for maintaining grid stability. This study proposes a machine learning approach to predict the Feasible Operating Region (FOR) of a medium voltage (MV) grid at its interconnection with a high voltage (HV) grid. The FOR serves as a method of aggregating the active and reactive power flexibilities (PQ-flexibilities) at the HV/MV interconnection while adhering to operational grid constraints. Knowledge of the underlying grid’s FOR is therefore essential for the effective operational management of the overlying grid, as it provides a precise representation of the permissible power exchange between the two grid levels. In this paper, the performance of both linear and nonlinear regression models is evaluated in terms of computational efficiency and accuracy, with all models benchmarked against the FORs derived using the Linear Optimal Power Flow (LOPF) method. Furthermore, a secure FOR is developed using a safety boundary approach, enhancing the study's applicability in real-world operational settings by ensuring that the predicted flexibility remains reliable under worst-case scenarios. The dataset, which is partitioned into training, validation, and test sets, is generated using Monte Carlo simulations combined with diverse contingency scenarios, thereby enhancing the model's robustness by capturing a wide range of realistic operating conditions. The results demonstrate a remarkable 185,000-fold improvement in computational efficiency, achieved without significant compromise in accuracy.