Automatic Power Quality Events Classifier based on hybrid CNN–LSTM network and multisine fitting algorithm

Power quality (PQ) disturbances can significantly affect the performance and longevity of electrical equipment, leading to system downtime, hardware degradation, and substantial economic losses. Real-time detection and accurate classification of such events are therefore critical to maintaining power system reliability. In this context, an Automatic Power Quality Event Classifier (APQEC) is crucial for the timely identification, segmentation, and classification of anomalies in electrical signals, enabling prompt grid maintenance and intervention. Many existing APQEC solutions proposed in the literature are based on centralised systems that are high-cost and therefore difficult to distribute, resulting in a low scalability monitoring system. This paper proposes the implementation of a PQ event classification framework based on a hybrid CNN–LSTM model combined with a multisinusoidal decomposition of the input signal. This framework can provide accurate PQ event classification in a distributed edge computing architecture under noisy conditions, combining decentralised detection with centralised classification in an edge computing architecture. A low-cost Local Distributed Node (LDN) is designed to be deployed across the monitored grid to detect PQ events and transmit compact features to a centralised Central Control Unit (CCU). The CCU receives the data from multiple LDNs and, for each of them, performs the classification. To permit the implementation of LDN with low-cost hardware, it is proposed the use of a Multisine Fitting Algorithm, which requires low computational capabilities and is therefore suitable for implementation on low-cost devices. This algorithm, which extracts the harmonic content required for classification, substantially reduces data exchange and system cost. At the CCU, a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model performs the classification and delivers robust performance under noisy conditions. The classifier is trained to recognise both single PQ events and composite event sequences for a total of 20 different classes. Numerical results demonstrate a classification accuracy exceeding 99% for single-event and event-combination classification within a signal-to-noise ratio (SNR) range of 10–30 dB, indicating the effectiveness of the proposed approach for scalable, real-world PQ monitoring.