Detecting Data Integrity Attacks on Correlated Solar Farms Using Multi-layer Data Driven Algorithm

This paper proposes a novel server-centric anomaly detection strategy for large-scale multi-modal time-series data from NSERC solar farm. The proposed framework utilizes spatial and temporal correlation between multiple solar farms to defend against data integrity attacks and learns the inter-farm/intra-farm relationship between measurements to perform anomaly detection. We demonstrated the performance of five machine learning algorithms—vector autoregressive models (VAR), deep neural networks (DNN), long short-term memory networks (LSTM), inverse PCA reconstruction (iPCA), and deep auto-encoders (DAE)—and evaluated the models on five varieties of data integrity attacks including replay, correlated, random, delay, and scaling attacks. Results show that the proposed framework is capable of detecting data anomalies over these large variety of attacks.