(374f) Optimal Storage Sizing of Sustainable Power-to-X System Using Generative Adversarial Networks

This paper proposes a data-driven framework to optimize the storage size and design of power-to-X (PtX) process incorporating the fluctuation of renewable power via the generative adversarial network (GAN). The main challenge in the PtX design is considering complex patterns of the renewable fluctuation and constraining potential usage of grid for carbon neutralization. Under the time-series uncertainty of renewable power, optimal storage and design should be figured out to minimize levelized cost of X (LCOX) while constraining grid usage in probabilistic domain. To address these challenges, The framework consists of three modules: renewable scenario generator, implicit scheduling model, and reliability analysis. The proposed framework employs a GAN as a renewable scenario generator to train temporal dynamics of renewable patterns in different geographical sites, using 20 years of historical data of wind and solar. The generator part of the GAN is used as a data-driven scenario generator that generates indistinguishable fake wind and solar profiles from noise vectors. In contrast to the model-based approach that uses empirical equations or probabilistic groups, the data-driven approach can generate an annual renewable profile simply by sampling a random vector from a Gaussian distribution. The implicit scheduling model determines the power allocation of renewable and grid power of the PtX systems, taking into account the mixing of solar and wind power, intermittent storage, and X production system. The time-series profile of power distribution can be computed using this model to satisfy continuous X production. The techno-economic performance of the PtX system is then computed using the determined allocation profile. The probabilistic distribution of LCOX and grid penetration can be directly computed under time-series uncertainty of the renewable, with the LCOX as the objective function and the grid penetration as the constraint. The reliability analysis is adopted to calculate the probability of grid penetration which violates constraints, and the subset simulation is employed in combination with the scenario generator and scheduling model to calculate the constraints. The overall framework has four deterministic inputs: installation capacity of wind and solar power plant, intermittent storage capacity, and X production capacity. The NSGA-ii optimizer iteratively updates the populations to determine the optimal population sets which minimize the probabilistic distribution of LCOX while satisfying grid penetration constraint. The paper includes a case study on the production of methanol and hydrogen on Jeju island in South Korea, describing how the proposed framework can provide reliable storage size and design, balancing the fluctuation of renewable power and power allocation integrated with the grid for sustainable production. The Pareto frontier shows a trade-off between the mean and variance in the LCOX distribution, with reducing the mean of LCOX increasing its variance, which describes the impact of uncertainty to PtX becoming more significant at lower LCOX distribution. The estimated minimal production cost for X is 1581.1 USD/ton, while the highest variance in cost under uncertainty is 4713.1. The mean cost increases to 2804.8 USD/ton with a minimal variance of 91.1. The Pareto curve line reveals two general trendlines: an increase in wind power plant installation capacity and a decrease in X production capacity. The wind fraction close to 1.0 results in maximum variance in production cost due to the lower fluctuations in wind power compared to solar power, making it uncertain. However, it is important to note that this trend is location-specific, as some studies report greater uncertainty associated with wind power in certain regions. Meanwhile, by reducing the production capacity, there is more room to accommodate fluctuations in renewable power. A case study involving the production of methanol on Jeju Island in South Korea demonstrates how the mean and variance of production cost and other performance indicators can be balanced in the reliable design of PtX processes under renewable energy uncertainty, providing insights for decision-making in such situations.