Accelerating optimal scheduling prediction in power system: A multi-faceted GAN-assisted prediction framework

This study introduces a comprehensive framework aimed at enhancing power system optimality through a two-stage optimization process and the development of a deep-based model for optimal scheduling prediction (OSP). Initially, a Bidirectional Long Short-term Memory (Bi-LSTM) architecture is employed to accurately forecast wind power in the first stage. Subsequently, a convolutional Generative Adversarial Network (GAN) model utilizes these predicted wind power values to generate synthetic scenarios. These scenarios, based on the preceding 10 days’ wind power predictions, serve as inputs for the subsequent power system optimization stage. To streamline computational efficiency, the power system optimization is conducted via a two-stage model. The outputs from this process, alongside other pertinent parameters, are utilized to train the proposed deep-based OSP model. The efficacy of the proposed model in rapidly and reliably predicting optimal scheduling is evaluated using the 118-bus power system. Results indicate that the innovative approach demonstrates exceptional speed and precision in determining optimal scheduling for the power system. Specifically, the proposed OSP model accurately forecasts optimal dispatch for ten days ahead in a mere 0.38 s, with an error rate below 0.001. Furthermore, the model exhibits a 92 % correlation in predicting optimal dispatched wind power. Sensitivity analysis highlights that optimizing the arrangement of the proposed deep-based model using an automatic hyperparameter optimization software framework (OPTUNA) can significantly enhance performance accuracy, potentially by up to 24 %.