Power quality assessment in solar-connected smart grids via hybrid attention-residual network for power quality (HARN-PQ)

Abstract

The aim of the proposed method is to solve the difficulties associated with anomaly detection and real-time data processing in complex network systems. The process begins by collecting data from internet of things (IoT) devices and smart grid sensors. Advanced interpolation techniques are used in pre-processing methods to deal with missing data, while the Isolation Forest algorithm is used to find outliers. Ensures data normalization through robust scaling, reducing the impact of outliers. Higher-order statistics such as skewness, kurtosis, and entropy measures, as well as various statistical metrics such as mean absolute deviation (MAD), interquartile range (IQR), and coefficient of variation (CV) are extracted in the feature extraction process. A unique method called hybrid horse-based zebra optimization (HHZO) is used to select features. It combines the zebra optimization algorithm (ZOA) and the horse herd optimization algorithm (HHO). Weighted ensemble energy quality residual attention network (WEARN-PQ) architecture is proposed for deep learning-based detection, which integrates extended recurrent neural networks (Stack-RNN) and stack-gated recurrent units (GRU) with attention layers and convolutional neural networks (CNN) with residual connections and attention mechanisms. To ensure reliability, split-sampling K-Fold cross-validation is used during training and validation.