Neural-Based Closed-Loop Framework for Pseudomeasurements Generation in Distribution System State Estimation

Abstract

Structural changes in modern power distribution networks are having an increasingly significant impact on monitoring and control applications, including distribution system state estimation (DSSE). These changes introduce uncertainties that degrade the quality of the available information, including measurements and pseudomeasurements (PMs), thereby limiting DSSE performance. With reference to PMs, which are nonmeasured information typically derived from a priori knowledge or historical data, machine learning (ML) techniques have emerged as a promising way to mitigate these effects. This article evaluates the effectiveness of using a neural-based approach to predict active and reactive power injections in a closed-loop framework that includes the PMG and the DSSE procedure. The performance achievable with neural models, such as linear neural networks, multilayer perceptron (MLP), recurrent neural network (RNN) architectures-including gated recurrent unit (GRU) layers and long short-term memory (LSTM) layers, and random forests (RFs), is compared. An assessment of weighting strategies for these predictions is also carried out to evaluate how these weights should be adjusted according to the prediction accuracy of DSSE-based neural models. The artificial neural networks (ANNs) were trained and characterized using real consumption data from the Forschungszentrum Jülich (FZJ) campus, and the influence of the improved PMs quality on the DSSE was evaluated via numerical simulations. Simulation results, based on historical data, prove the validity of the proposed approach as a reliable monitoring tool for system operators.