Data anomaly detection in photovoltaic power time-series via unsupervised deep learning with insufficient information

Anomaly detection in photovoltaic (PV) systems is essential to improving reliability, ensuring electricity production and equipment safety, and decreasing their negative impact on the economy of the operation system. In many real-world scenarios—such as limited historical data, incomplete documentation, varying conditions, data corruption, or privacy issues—insufficient and unlabelled data challenge traditional anomaly detection and supervised learning methods for PV systems. Therefore, this paper proposes an effective unsupervised data anomaly detection model based on a deep neural network autoencoder. This model does not require prior knowledge about the system and accurately identifies PV system anomalies with limited information. The proposed model only uses measured PV power production as input and does not need additional information on PV system parameters or measurement data. Additionally, we derived an optimal threshold to detect anomalies based on the mean and standard deviation of the reconstruction error, resulting in a significant improvement in the F1-score from 0.9123 with the traditional approach to 0.9993. Lastly, a novel locally adaptive mechanism based on Dynamic Time Warping (DTW) error analysis is proposed to effectively locate anomaly segments by considering the shape of anomalous parts within the input time series data. The proposed model is validated on a real PV power plant in Genoa, Italy. The case study results demonstrate that the model outperforms other unsupervised machine learning models with a 0.9538 F1-score in testing and shows performance comparable to that of advanced supervised models, including XGBoost and deep neural networks.