Time-Series Anomaly Detection in Off-Grid Solar-Powered Security Systems Using Deep Learning Approaches

Abstract

Reliable operation of off-grid solar-powered security systems is essential for maintaining uninterrupted surveillance and monitoring in remote locations. This study proposes a fully unsupervised deep learning framework for detecting energy-related anomalies in multivariate solar-system time-series data before they affect system operation. To this end, seven deep learning-based anomaly detection models, namely, LSTM, GRU, Transformer Autoencoder, CNN-LSTM, CNN-GRU, LSTM Autoencoder, and GRU Autoencoder, were developed and evaluated using both a real-world solar energy dataset and an open-source Kaggle dataset. All models were trained in an unsupervised manner without labeled anomaly data. Anomalies were identified using prediction or reconstruction errors combined with percentile-based thresholding, while model performance was evaluated using MSE, RMSE, and MAE metrics. The results showed that all models were able to learn temporal patterns and identify statistically unusual operating conditions. The Transformer Autoencoder achieved the lowest MSE and RMSE values, whereas the GRU model yielded the lowest MAE. SHAP-based explainability analysis was further employed to interpret model decisions and identify the most influential variables contributing to anomaly detection. Experiments on both datasets provided evidence that the evaluated approaches can identify deviations from learned temporal patterns under different data characteristics.