Hybrid binarized neural network for high-accuracy classification of power quality disturbances

Abstract

Binarized Neural Network (BNN) is a technique for reducing computational complexity and memory requirements by constraining weights and activations to binary values, enabling deployment on lightweight platforms. However, the current BNNs confront a problem of limited accuracy due to significant information loss, thereby failing to deal with in complex tasks, especially in power quality disturbance (PQD) classification. To solve this problem, we propose a hybrid binarized neural network (HBNN) model that reintroduces full-precision convolutional layers. This allows for the retention of more details and features from the original data, thereby enhancing the network’s representation of the data. HBNN enhances the nonlinear expressive capability by incorporating a full-precision convolutional layer as the input layer, while the subsequent layers maintain the binarized layer to reduce model complexity, enabling the network to better adapt to lightweight platforms. We validate the proposed method and the alternative baselines for classifying 16 types of power quality disturbances. Experiments demonstrate that HBNN improves accuracy by 9.13% compared to BNN.