JISE

Industrial end-users always demand for better power quality to retain grid’s efficiency and maintain machinery health. However, increasing complexity of multiple energy systems increase the risk of deterioration in the quality of power supplies. An effective power quality disturbances detection tool is highly needed to automatically detect the unusual event exhibiting inside the systems. Power quality disturbances can be categorised into two category, i.e. (a) short and fast transient disturbances and (b) long and slow disturbances, which complicates the process of classification using similar model. In this paper, a hybrid deep learning model consisting of multi-resolution transformation and deep neural network is proposed for power quality disturbances detection. Transformer network has been proposed to improve the classification and computation performance with its multi-head attention mechanism and parallel computing characteristics. The proposed hybrid model first transforms input signal into multiple frequency components using multi-level signal decomposition signal via wavelet transform. A layer of convolutional kernel is used to obtain the spatial and temporal features from the wavelet components. The process is followed by higher order latent feature extraction using transformer network which includes a layer of transformer encoder and a pooling mechanism. The proposed model is able to outperform other deep neural network models with better accuracy despite of the noisy condition.