Physics-Informed Domain Adaptation for Electrical Energy Monitoring

Changing grid conditions challenge modeling assumptions for power systems. For example, electrical signatures may differ between “islanded” and “grid-connected” modes of a microgrid’s operation. Unmodeled variance and shifts in data distributions confound data-driven techniques that model practical physical systems. With proper preprocessing, knowledge of the system’s underlying physics can guide the training and use of data-driven models of load behavior to provide domain adaptation that “undoes” these effects. These adaptation techniques provide greater feature space separability, enabling better visualization, clustering, and load identification. This work presents a general technique for deploying physics-informed domain adaptation on monitoring systems for microgrids with evolving voltage conditions. Case studies across multiple types of electrical loads demonstrate that these techniques can remove significant standard deviation in electrical features, ranging here from a 37.8% to 89% reduction.