Anomaly Detection Model Development

After identifying the normal operating modes of 2020 Fall chillers, we developed models with the normal operating patterns. Predictions are then generated at each time step and the errors in Predictions represent deviations from expected behavior. A reconstrunction error above a pre-chosen threshold means an anomaly is found. We evaluate the model trained on 2020 Fall data on 3 different datasets: (1) 2020 Fall normal operating patterns with outliers, (2) The whole set of 2020 Fall data, (3) 2021 Fall data

Autoencoder trained on 2020 Fall chillers normal operating patterns

LSTM-based autoencoder trained on 2020 Fall chillers normal operating patterns