State Estimation in Power Distribution Grids Using Deep Unfolding

The power distribution network has undergone significant changes in recent years, primarily due to the integration of renewable energy sources and electric vehicles. These changes have led to new challenges in distribution grid management, which relies on situational awareness to ensure safe, reliable, and secure operations. Real-time distribution system state estimation (DSSE) is essential for achieving this situational awareness. Previously proposed DSSE methods are computationally expensive and lack scalability. Therefore, recently, researchers have turned to learning-based approaches such as deep neural networks (DNNs) to overcome these limitations. However, DNNs are susceptible to over-fitting. To address these challenges, this paper proposes a novel model-based neural network approach, called deep unfolding, for fast and accurate estimation of system states in a low observable distribution network. Unlike model-agnostic neural networks (NNs), which are difficult to tune and train, the proposed model-based NN is created by "unfolding" the alternating direction method of multipliers (ADMM) solver. This article proposes two methods for state estimation: a purely data-driven approach called model-free DSSE network (MFDSSE-net) and a model-based DSSE network (MBDSSE-net) that incorporates network information by including the linearized power flow constraints into the optimization framework. The simulation results are compared with the state-of-the-art ADMM method and weighted least square method and demonstrate the efficacy of the proposed approaches in the IEEE-123 bus and a 559 bus network. Simulation results show accurate state estimation is possible with nearly 300% reduction in simulation time when deep unfolding is used.