Point and contextual anomaly detection in building load profiles of a university campus
Long Wang, Marian Turowski, Meng Zhang, Till Riedel, Michael Beigl, Ralf Mikut, Veit Hagenmeyer · 2020
The increasing use of smart meters enables the monitoring and diagnostics of underlying systems. The application of data analysis methods can help to automate monitoring and diagnostics such that human intervention is limited to the situations where and when it is necessary. In a smart grid, diagnostics can relate to faulty smart meters and unusual consumption, corresponding to point anomalies and contextual anomalies. This work compares a Deep Neural Network Regression, an Autoencoder with reconstruction, and the encoder of the Autoencoder as anomaly detection methods. The three models are evaluated on real-world building load profiles of a university campus containing such anomalies. The results demonstrate that the proposed models have superior detection accuracies over benchmarks and differ in their discrimination between anomalies and normal electrical load profiles. At the same time, the models correctly identify different anomalous electrical load profiles that were wrongly labeled as normal.