The system frequency is a critical measure of power system stability and understanding, and modeling it are key to ensure reliable power system operations. Koopman-based autoencoders are effective at approximating complex nonlinear data patterns, with potential applications in the frequency dynamics of power systems. However, their non-invertibility can result in a distorted latent representation, leading to significant prediction errors. Invertible neural networks (INNs) in combination with the Koopman operator framework provide a promising approach to address these limitations. In this study, we analyze different INN architectures and train them on simulation datasets. We further apply extensions to the networks to address inherent limitations of INNs and evaluate their impact. We find that coupling-layer INNs achieve the best performance when used in isolation. In addition, we demonstrate that hybrid approaches can improve the performance when combined with suitable INNs, while reducing the generalization capabilities in combination with disadvantageous architectures. Overall, our results provide a clearer overview of how architectural choices influence INN performance, offering guidance for selecting and designing INNs for modeling power system frequency dynamics.