Reliable Power System Operations with Renewable Ramps and Asymmetric Faults using Graphical Deep Learning

The growing integration of renewable sources, particularly wind energy, into modern electricity grids poses new operational challenges due to their inherently intermittent and stochastic nature. For places with high wind energy integration, rapid changes in generation can lead to significant supply-demand discrepancies, hence the risk of compromising reliability for systems. Large-scale ramp events, where generation can vary by several gigawatts (GW) over a short period (1 hour), are commonly occurring and hence emphasize how accurate and timely forecasting tools are needed. Additionally, increased deployment of distributed energy sources (DERs) and inverter-based systems has increased the sensitivity of power systems to asymmetric faults, some of which are single-line-to-ground, line-to-line, and double-line-to-ground faults. If not identified quickly, faults of these types can rapidly spread through the grids, having the potential to lead to equipment damage, load interruption, or large-scale system instability. This work studies the application of state-of-the-art graph neural network (GNN) architectures to two key challenges of power grid management: (i) predicting swift wind energy production ramps, and (ii) localizing unbalanced faults based on phasor measurement data. By capitalizing on electrical grid's intrinsic graph structure, GNNs can discover rich spatio-temporal correlations of complex patterns missed by traditional forecasting and fault localization schemes. For wind ramp prediction, the model integrates meteorological data and generates data across multiple sites to detect patterns between far-separated sites, boosting accuracy of predictions regarding abrupt changes of energy output. Concerning identification of asymmetric faults, the technique employs measurements of voltages and current across multiple buses to describe directional impacts and space dynamics, supporting precise identification of fault type and their sites. With this work, it is illustrated that GNN-based schemes surpass traditional deep learning techniques by efficiently incorporating both renewable generation's and system measurement's spatial and temporal information. The framework offers improved situational awareness for system operators, supporting increased reliable and resilient operation of grids having high renewable integration. Through addressing both asymmetrical fault vulnerability and generation variability, contributions are made towards more intelligent and adaptable power systems designed for a quickly changing energy landscape's demands.