Warped Gaussian Processes for Power Grid Data Analysis

Reliable power grid operation is critical for supporting nearly every aspect of modern life. Outages due to extreme weather, equipment malfunction, or even cyber attacks can disrupt infrastructure and create serious, lasting problems for communication and emergency services. The Grid Communications and Security group at Oak Ridge National Laboratory (ORNL) works with industry partners to develop the Autonomous Intelligence Measurements and Sensor Systems (AIMS) project. AIMS is a system using a machine-controlled response team with sensors, mobile platforms, communications and data storage and analysis components. When it comes to monitoring the power grid, leveraging drone and sensing technology provides new opportunities to improve operations. The different data streams comprise an interrelated, complex system with significant amounts of data points. All the data collected by the system will need to be analyzed in order to transform it into useful information. A Gaussian process (GP) is a probabilistic method that has been successfully applied to grid data in the past. However, the assumption that the data and its noise follows a Gaussian distribution is not always realistic. Warped Gaussian Processes (WGP), a generalization of a GP, retains the power of a GP but is not reliant on that assumption. Given the non-stationary and non-linear data that tends to accompany the power grid, WGP is an underutilized method with potential to improve grid modeling. A Pandapower IEEE 14-Bus system was used to generate power grid data to evaluate the performance of WGP as a method for data analysis in the grid context. Other data include real-world signal data from the Grid Event Signature Library and weather-load data from Panama's national grid operator, CND. This analysis demonstrates that WGP is suitable for power grid data analysis because it improves upon GP and other traditional statistical methods in terms of parameter estimation, signal filtering, and load forecasting. Future work should focus on managing the computational complexity of this method and integrating artificial intelligence and other machine learning techniques into data analysis pipelines.