An optimized generative adversarial network for state estimator in electric power systems

• The well-trained model can rapidly produce state estimation results without iterative process. • Generative network exhibits low sensitivity to bad data used in state estimators. • Topology of generative adversarial network is tuned by particle swarm optimization. Energy management system (EMS) relies on complete information based on measured data in an electric power system to provide useful decision making for dispatchers. State estimation is a fundamental tool employed in an EMS to determine the states, including voltage magnitudes and phase angles, among others, using an appropriate set of measurements. Traditional state estimation methods typically rely on criteria like the maximum likelihood criterion, weighted least-squares criterion, or Kalman filter to compute these states. However, these conventional methods may be susceptible to issues such as noise, false readings, or missing measurements. To address these challenges, an innovative state estimator based on a Wasserstein Generative Adversarial Network (WGAN) is proposed in this paper, implemented through two iterative optimization loops. In this approach, the data distribution is captured by the generative network, while the probability that sampled datasets originate from the training datasets is evaluated by the discriminative network. Notably, the structure parameters and hyperparameters of both networks are determined using particle swarm optimization, thereby avoiding heuristic selection. With an appropriately designed network architecture and training strategy, the proposed WGAN is capable of performing effectively even when limited datasets are available. The effectiveness of this approach is demonstrated using the IEEE 30- and 118-bus test systems, particularly when dealing with problematic data scenarios, such as missing data, reversed data, and large variances. In the proposed method, only the scenario with three missing data points results in measurement residuals exceeding the threshold value (1313.80) in the IEEE 118-bus system. All other types of bad data do not cause the proposed method to produce high measurement residuals, demonstrating its robustness and insensitivity to bad data. In contrast, all measurement residuals involving three types of bad data obtained using the traditional weighted least-square criterion exceed this threshold.