Predicting Minute-Level Power Grid Conditions Using Generative Adversarial Networks

As renewable energy systems are increasingly integrated into grid operations, optimal and flexible use of nuclear energy resources in such integrated setups requires accurate forecasting of grid demand and power produced by renewable energy systems. Precise prediction of contributions from renewable energy sources is vital for planning nuclear-provided generation in a future system where nuclear energy is expected to be operated flexibly. This becomes highly crucial for ensuring power grid stability in the face of increasing power demand, most notably from recent and upcoming high electricity consumption from data centers. This study investigated generative adversarial networks (GANs) to predict minute-level power grid data, focusing on solar power, wind power, and load power in California Independent System Operator (CAISO) Zone 1. Conditional GAN (cGAN), Wasserstein GAN (WGAN), Gated Recurrent Unit (GRU), and sequence-to-sequence (seq2seq) GRU were used to capture the complex temporal dependencies in the data. The study systematically evaluated model performance using various lookback (1 h, 12 h) and lookforward (1 h, 12 h) windows, assessing each model's accuracy using metrics such as the root mean square error and the mean absolute error. The results demonstrate that WGAN achieves higher accuracy than cGAN, GRU, and seq2seq GRU in almost all scenarios. This higher accuracy is attributable to WGAN's use of the Wasserstein distance as the loss function, which enhances stability and mitigates mode collapse. The results in this paper show the ability of generative models to generate accurate time series data that may be used to assist in optimizing nuclear energy generation in future grids.