A multi-modal feature combination mechanism for identification of harmonic load in distribution networks based on artificial intelligence models

Abstract

The challenge of harmonic pollution in active distribution networks is inherently complex, a data-driven approach being necessitated to comprehensively capture the nonlinear and non-stationary characteristics of harmonic power sequence signals, thereby enhancing recognition accuracy. To achieve intelligent identification of harmonic loads within distribution networks, an innovative methodology that integrates parameter-optimized variational mode decomposition with sequential neural networks is proposed. Firstly, based on IEEE Std. 1459-2010 power theory, the harmonic apparent power distortion caused by nonlinear loads is calculated. Secondly, using an optimization algorithm, the penalty parameter and the number of intrinsic mode functions in variational mode decomposition are fine-tuned to decompose the harmonic power sequence and extract intrinsic mode functions. The most suitable intrinsic mode sequences are selected as input features for sequential neural networks training. Finally, a multi-modal feature tensor combination mechanism that integrates reshaped vector layers into the sequential neural networks architecture is introduced, enabling adaptive extraction of spatial–temporal characteristics and significantly improving the accuracy of harmonic load identification without prior knowledge of their spectral features.