A multimodal deep learning model with differential evolution-based optimized features for classification of power quality disturbances

• Published in: Journal of Electrical Systems and Information Technology Vol. 12; no. 1; pp. 11 - 21
• Main Author: Islam, Md Nurul
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.12.2025; Springer Nature B.V; SpringerOpen
• Subjects: Accuracy; Artificial Intelligence; Artificial neural networks; Classification; Clean energy; Computing time; Data science; Deep learning; Deep neural network; Differential evolution; Discrete wavelet transform; Distributed generation; Disturbances; Electrical Engineering; Electrical loads; Electrical Machines and Networks; Engineering; Evolutionary computation; Fast Fourier transform; Fourier transforms; Fuzzy logic; Genetic algorithms; Information Systems and Communication Service; Machine learning; Mechanical Engineering; Multimodal feature fusion; Neural networks; Optimization; Power Electronics; Power management; Power sources; Signal processing; Signal quality; Support vector machines; Voltage sags; Wavelet transforms; Differential evolution; Deep neural network; Discrete wavelet transform; Multimodal feature fusion; Fast Fourier transform
• ISSN: 2314-7172; 2314-7172
• DOI: 10.1186/s43067-025-00194-0

Electric power quality disturbance is a grave concern, as renewable sources are increasingly connected to the microgrid to provide green energy. The use of power electronic devices, intermittent nature of renewable power sources and sudden electrical load changes are the primary contributors to power quality degradation that eventually incorporates voltage sag, swell, interruption, transient, oscillation, harmonics, flicker and combination of two or more of them to the pure sinusoidal voltage supply. Detection and classification of the different power quality disturbances has become one of the current priority research topics for last few years. Fourier transform and wavelet transform are the signal processing tools used to extract wide range of frequency and time-domain features of the anomalies of the disturbance signals before being processed by the deep learning model. Differential evolution is used as an optimization tool to select the most relevant features and reduce the computational time. The selective important features are given to a multimodal deep learning model built using multiple layers of convolution neural network, long short-term memory and deep neural network to extract and infuse the deep spatial and temporal features of the signals before being fed to the SoftMax neural network for classification task. The work shows a great accuracy of 99.6%, even with the big dataset of 121,000 samples and 20 types of power quality disturbance signals.