Physics-guided Deep Learning for Distribution Network SE

With the rapid advancement of modern power systems, the structure and operation of the grid have become increasingly complex, demanding higher levels of real-time accuracy and efficiency in state estimation (SE). To address the limitations of traditional SE methods for distribution networks, which often rely heavily on model precision, exhibit lower estimation accuracy, and have slower response times, a novel approach called Physics-Guided Temporal Convolutional Network (PGTCN) is proposed. The PGTCN method combines the strengths of data-driven techniques with the physical advantages of model-driven approaches. It begins by training a Temporal Convolutional Network (TCN) model using historical operational data from the distribution network. The model's predictions of the state variables are then integrated into the power flow equations, allowing for the examination of their consistency with the underlying physical relationships. Through this process, the PGTCN model achieves synergy between data-driven and model-driven methodologies, resulting in higher accuracy and improved estimation precision. Simulations conducted on a three-phase unbalanced distribution system with IEEE 14 nodes demonstrate the superior performance of the proposed SE method. Overall, the PGTCN approach offers a promising solution to the challenges posed by the ever-evolving landscape of modern power systems, promptly ensuring robust and accurate SE.