Fast Deep-Learning-Based Recognition of Multiple Power Quality Events Under Noise and DC Offset

The wide integration of renewables, e.g., photo-voltaics, into the power grid can result in decreased power quality. Real-time recognition and classification of Power Quality Events (PQEs) are of great interest to the power system operators, to maintain an acceptable quality of the delivered power across the grid. This paper investigates the potential of using learningbased algorithms to obtain a single solution that can accurately (i) recognize and classify both single and multiple simultaneous PQEs, (ii) in a timely manner, and (iii) under practical sources of measurement error such as noise and dc offset. Simulations are carried out in two steps. Firstly, the performances of 31 reputable learning-based algorithms are evaluated, in MATLAB, to demonstrate the effect of the aforementioned signal variations and sources of error on the accuracy of PQEs recognition and classification. This effect is also verified using the advanced TDistributed Stochastic Neighbor Embedding algorithm. Afterward, a Convolutional Neural Network (CNN) is implemented to recognize and classify PQEs under the aforementioned factors. Our results show that, despite the given problem's complexity, neural-network-based techniques are able to achieve higher accuracy than the other studied techniques. CNN can achieve 95.5% accuracy.