Privacy-Preserving Utilization of Distribution System Flexibility for Enhanced TSO-DSO Interoperability: A Novel Machine Learning-Based Optimal Power Flow Approach

TL;DR

This paper tackles the privacy challenge in TSO-DSO interoperability by introducing an ML-based framework that encodes distribution-system constraints as non-sensitive models, enabling the TSO to solve a privacy-preserving AC-OPF in a single round and directly dispatch DGs. A novel NN-guided polytope represents DS feasible regions, while quadratic regressors map DS variables to PCC injections; dataset generation via Latin Hypercube Sampling supports arbitrary convex PQ characteristics beyond rectangles. The approach is extended to multiple DSs and multiple PCCs, with demonstrated accuracy and negligible communication overhead, achieving comparable feasibility to standard AC-OPF and enabling secure, scalable DS flexibility utilization. The results highlight significant practical impact for reliable, privacy-protecting TSO-DSO coordination in meshed networks with diverse FPUs. The proposed method thus enhances interoperability and operational efficiency without exposing sensitive DS data.

Abstract

Due to the transformation of the power system, the effective use of flexibility from the distribution system (DS) is becoming crucial for efficient network management. Leveraging this flexibility requires interoperability among stakeholders, including Transmission System Operators (TSOs) and Distribution System Operators (DSOs). However, data privacy concerns among stakeholders present significant challenges for utilizing this flexibility effectively. To address these challenges, we propose a machine learning (ML)-based method in which the technical constraints of the DSs are represented by ML models trained exclusively on non-sensitive data. Using these models, the TSO can solve the optimal power flow (OPF) problem and directly determine the dispatch of flexibility-providing units (FPUs), in our case, distributed generators (DGs), in a single round of communication. To achieve this, we introduce a novel neural network (NN) architecture specifically designed to efficiently represent the feasible region of the DSs, ensuring computational effectiveness. Furthermore, we incorporate various PQ charts rather than idealized ones, demonstrating that the proposed method is adaptable to a wide range of FPU characteristics. To assess the effectiveness of the proposed method, we benchmark it against the standard AC-OPF on multiple DSs with meshed connections and multiple points of common coupling (PCCs) with varying voltage magnitudes. The numerical results indicate that the proposed method achieves performant results while prioritizing data privacy. Additionally, since this method directly determines the dispatch of FPUs, it eliminates the need for an additional disaggregation step. By representing the DSs technical constraints through ML models trained exclusively on non-sensitive data, the transfer of sensitive information between stakeholders is prevented.

Figures (8)

• Figure 1: Schematic representation of the proposed method.
• Figure 2: Data sampling approach using LHS.
• Figure 3: Schematic representation of the polytope.
• Figure 4: The architecture of the novel tailored NN.
• Figure 5: Single line diagram of the integrated power system.