Attention Mechanism for Increasing the Performance of LSTM Model on Anomaly Detection in Photovoltaic Power Generation

Anomalies can occur in photovoltaic (PV) power generation when there is a deviation from the expected performance, which can cause damage and reduce the life of the equipment involved. However, state-of-the-art studies have not explored contextual anomalies and collective anomalies caused by weather and PV power generation equipment conditions. This study aims to improve anomaly detection performance in PV power generation using the long short-term memory (LSTM) model with the attention layer. First, we use the PV power generation anomaly dataset from Zenodo. Then, we apply mean decrease in importance (MDI) feature importance to select the most correlated features with anomalies. We test the performance of the attention layer in improving the performance of LSTM to detect anomalies with four metrics: accuracy, precision, recall, and f1-score. Finally, we test the proposed model's performance in detecting three types of anomalies: Point, contextual, and collective anomalies, using the confusion matrix and sensitivity score of each label. The test results show that with MDI feature importance, we receive six features that provide optimum accuracy in our anomaly detection model, which is 0.978. Then, applying the attention layer to the LSTM model increases the accuracy, precision, recall, and f1-score of the model, which is from 0.958, 0.933, 0.952, and 0.941 to 0.997, 0.995, 0.998, and 0.996, respectively. Finally, the confusion matrix shows that our proposed LSTM+attention layer model can detect point, contextual, and collective anomalies with sensitivity values of 0.997, 1.000, and 0.988, respectively.