Neural Networks to Improve Distribution State Estimation—Volt Var Control Performances

Distribution systems must be ready to face upcoming technical and economic constraints: increase of Distributed Generation (DG) connections, changes in network losses and voltage profiles, among others. In this context, new centralized automation functions in distribution system control centers are needed in order to ensure the control of both distribution network and connected DGs. Consequently, state estimators need to be developed for future distribution systems to assess the network´s state in real time, i.e., 10 minutes typical time frame, based on real, pseudo-, and virtual measurements. Such state estimation functions are widely used for transmission systems but cannot be transposed directly into distribution systems. Indeed, one of the main issues is the lack of sensors in the distribution network, requiring additional load models to solve observability issues. These load models (also called pseudomeasurements) are usually active and reactive power models at medium to low voltage (MV/LV) substations using often very inaccurate information from historical database or other estimated load curves, for instance. The scale of these errors makes the estimation of all variables in the distribution network difficult. This paper proposes a pseudomeasurement estimation using neural networks in order to improve the results of a distribution state estimator (DSE), used as inputs to a centralized Volt and Var control function.