A novel hyperparameter tuned deep learning model for power quality disturbance prediction in microgrids with attention based feature learning mechanism

Abstract

Microgrids (MGs) have become a reliable power source for supplying energy to rural areas in a secure, consistent, and low-carbon emission manner. Power quality disturbance (PQD) is a common issue that reduces the MGs networks’ reliability and restricts its usage on a small scale. The performance, reliability and lifetime of the various power devices can be affected due to the problem of PQD in the network. Researchers have proposed numerous PQD monitoring techniques based on artificial intelligence. However, they are limited to low margins and accuracy. So, this paper suggests a novel hyperparameter-tuned or optimized deep learning model with an attention-based feature learning mechanism for PQD prediction. The critical stages of the proposed work, such as data collection, feature extraction, and PQD prediction, are as follows. The PQD signals are first produced using the IEEE 1159 standard. Following that, the original time-domain features are directly recovered from the dataset, and the frequency-domain features using discrete wavelet transform (DWT). The extracted features were fed into visual geometry group 16 with multi-head attention and optimal hyperparameter-based bidirectional long short-term memory (V16MHA-OHBM) to perform spatial and temporal feature extraction. These extracted features are concatenated and given to the fully connected layer to forecast the PQD. The results showed that the suggested approach surpasses the prior state-of-the-art algorithms when trained and tested using 16 different types of synthetic noise PQD data produced using mathematical models in line with IEEE 1159.