Phase-angle-free harmonic coupling analysis and injection sites identification approach via data-driven regression model of harmonic voltage versus current

Abstract

Harmonic coupling analysis and injection site identification are essential for maintaining reliable power system operation. Conventional approaches rely on detailed system models and synchronised multi-site phase-angle measurements, which are seldom publicly available, and deploying such metering network-wide is impractical. This paper introduces a data-driven approach for harmonic coupling analysis and injection site identification, maintaining high reliability while requiring only limited measurements. The method involves two main steps. First, a Multi-Compression Refined Self-Attention Network (MCReSANet) is used to model the relationship between harmonic voltages and currents in low-voltage (LV) grids. This model does not require phase angle information and supports both deterministic and probabilistic analyses. Second, SHapley Additive exPlanations (SHAP) values are applied to interpret the trained regression model, enabling qualitative assessment of correlation strengths across different harmonic components. The method is validated using two real-world LV datasets. Compared to benchmark models (Convolutional Neural Network (CNN) and Multi-Layer Perceptron (MLP)), the MCReSANet-based model improves accuracy by 10%–20% in both deterministic and probabilistic analysis. In addition, SHAP-based harmonic coupling and injection site analysis using MCReSANet shows more stable and interpretable results with lower noise levels than CNN and MLP, across both single and multiple site applications.