A semi-supervised anomaly detection approach applied to solar energy generation

As the power grid grows and becomes more efficient, the need for better integration with renewable energy generation data also progressively increases. In this sense, about solar energy sources, it becomes important, for example, to detect anomalies robustly, especially when observing the susceptibility of power generation through measurement equipment, weather data, and even measurement failures. These failures can compromise the accuracy of predictions in the power distribution network and thus accentuate the urgency of effective methodologies for detecting such anomalies. Thus, this study presents a semi-supervised anomaly detection framework built for real solar power generation data. As an alternative to unsupervised frameworks, the suggested method uses typical solar power generation patterns, using manually labeled data to train the models exclusively on health points, minimizing the training runtime and improving the anomaly detection results. Meteorological data is also used. Anomaly detection is evaluated using accuracy, precision, recall, area under the curve, and F1 score, with prediction intervals at 70%, 80%, and 90% levels. The results demonstrated that kNN (k-Nearest Neighbors) and HBOS (Histogram-Based Outlier Score) models outperformed other models, and the execution time of semi-supervised models was reduced by 60% when compared to unsupervised frameworks.