Transformer-based architectures for anomaly detection in solar-wind integrated energy infrastructures

The growing incorporation of variable solar and wind power in electrical grids leads to serious issues concerning reliability due to the complexity of spatial and temporal behavior caused by factors like variations in environmental conditions, degradation of equipment, failures of sensors, and exceptional operating conditions. Therefore, we develop an innovative transformer-based deep-learning approach for anomaly detection in integrated solar-wind energy systems using real-life multivariate time series datasets from a renewable energy facility situated in Kayathar, Tamil Nadu, India. The framework proposed in this paper consists of data preparation and normalization, sliding window conducted as a set of sequences, anomaly data creation based on domain knowledge, transformer encoder enhanced with multi-head self-attention technology and the attention mechanism-based data interpretation. The effectiveness of the framework is compared with LSTM, GRU and CNN-LSTM using MSE, accuracy and ROC-AUC. The obtained results indicate the sufficient performance of the transformer model with the lowest MSE value of 0.00641 in comparison to 0.00831, 0.00859 and 0.00792 achieved for LSTM, GRU and CNN-LSTM respectively. The transformer model has also achieved the value of accuracy equal to 95% together with ROC-AUC of 0.89. However, at the default classification threshold, this accuracy was driven almost entirely by the majority (normal) class, and the model failed to identify any of the 10 labelled anomalies in the confusion-matrix evaluation (recall = 0% for the anomaly class), reflecting a severe 18.8:1 class imbalance in the test set. The correlation and attention analysis performed in the study shows that the model is capable of revealing significant interdependence among various environmental parameters and renewable power generation rates such as solar radiation and solar energy production ( r = 0.84) and wind speed and wind energy production ( r = 0.90).