Quantifying Distribution Leakage in Wind Power Prediction Models 108 views

Rapid growth in wind energy generation promises significant economic and environ- mental benefits, but realizing these gains depends on accurate power‐prediction mod- els trained on large, diverse datasets. However, concerns over proprietary Supervisory Control and Data Acquisition (SCADA) system data currently limit industry‐wide data sharing. This work quantifies the privacy risks of sharing wind‐turbine SCADA system data by measuring information leakage from single‐layer Long Short-Term Memory (LSTM) power‐prediction models via distribution‐inference attacks. Using eight years of open-source SCADA system data, we extract fixed-length windows of six continuous, operator-sensitive features and train LSTMs under three realistic information-sharing settings. We then evaluate distribution inference risk using the Loss Test and a modified Threshold Test, which leverages adversary-trained shadow models and mean squared error thresholds to distinguish between datasets that differ only in the proportion of a sensitive attribute, in our regression setting. Our experi- ments reveal that wind‐speed–derived properties leak the most information, whereas more complex features—turbulence, misalignment power loss, time availability, and capacity factors—leak less. While attack effectiveness varies with the underlying wind‐speed distribution for different sensitive attributes, it remains largely unaffected by the choice of data‐partitioning strategy.

MS (Master of Science)

Long Short-Term Memory; Distribution Inference

English

2025-07-22