Distribution system state estimation using physics-guided deep learning approach

In the modern distribution management system, state estimation (SE) has a key role in monitoring the entire distribution system using advanced measurement units. Since the conventional weighted least squares (WLS) approach faces convergence issues due to insufficient measurement devices in the distribution system, deep learning (DL) models are highly effective in performing SE. However, the existing DL models suffer from generalizability issues, unawareness of the system’s physics, and scalability issues. To address this, physics-based deep learning models have been introduced. In this paper, a physics-based temporal convolutional neural network (Ph-TCN) is proposed to perform distribution system SE (DSSE). The first stage of the proposed approach utilizes a conventional TCN model in a supervised manner for the estimation of system states, while the latter stage incorporates physics-based properties to reconstruct the measurements. The use of the TCN model in first stage facilitates the capturing of temporal features with a large receptive field, whereas the inclusion of physics equations in the second stage with problem-specific Huber-loss function enhances the performance of the model to provide accurate SE results. Simulation studies are performed on modified IEEE 13-node and IEEE 37-node distribution test systems and the corresponding results have been compared with the WLS approach and existing DL models. The numerical results show that the proposed Ph-TCN approach has exhibited superior performance and its robustness is also tested under the presence of different percentages of missing measurement data. Finally, the models’ performance is also evaluated for the modified IEEE 123-node distribution system to check for the scalability issues. •A Physics-based TCN model is developed for distribution system state estimation.•A novel loss function by combining MSE and problem-specific Huber loss is used.•The model performance is compared with WLS, physics-based deep learning approaches.•The robustness of the proposed model is tested at different levels of missing data.•The results prove that the proposed model performance is effective for DSSE.