LSTM Network-Based Scheme for Automatic Characterization of Power Quality Disturbances

Recognition of power quality (PQ) troubles is a critical task in the electrical power industry. Most previous works solve the classification problem using separate feature extraction phase and classification phase. Each phase has its own techniques, and consumes a computation time. This study proposes to utilize the long short-term memory (LSTM) network as a deep learning model to classify the PQ events in one shot. The LSTM network uses its particular processing to classify a PQ event signal directly by reading its time-sequence data. Then, a dedicated post-classification algorithm (PCA) extracts start time, end time, duration, amplitude, and total harmonic distortion from the disturbance signal. Ten classes of simple and combined PQ events including interruption, sag, swell, surge, sag plus harmonics, swell plus harmonics, and composite disturbance are catered. A large data base of 1000 PQ disturbances for training, validating and testing the LSTM network is generated in MATLAB programming environment. The scheme is evaluated by 200 various test cases, which are classified with 100% precision for signal to noise ratio above 40 dbs. The proposed scheme shows excellent performance in automatic insightful description of PQ events by comparison to literature.