Long-horizon forecasting in nuclear power plants is essential for operator support under abnormal and accident conditions, where timely understanding of future parameter evolution can improve decision-making and help maintain safety margins. In such situations, operator support systems should not only interpret the current plant state but also anticipate the future trends of multiple key variables over extended time horizons. However, as the forecasting horizon becomes longer and the number of variables increases, direct forecasting models often suffer from degraded performance due to high-dimensional correlations and nonlinear dynamic behavior. To address this challenge, this study investigates a latent prediction framework for long-horizon forecasting and focuses on optimizing the latent dimension, which is a key design variable in latent-space modeling.

A pilot study on nuclear power plant emergency scenario forecasting showed that a two-stage latent-space approach, in which an autoencoder compresses future parameter trends into latent vectors and a predictor estimates those latent vectors before reconstruction through a decoder, can outperform direct forecasting models in both accuracy and efficiency for long-horizon prediction. Building on this motivation, the present study treats the latent dimension not as a fixed empirical parameter but as an optimization target. Multiple latent dimension candidates were evaluated by jointly considering reconstruction quality and forecasting performance.

The results indicate that an excessively small latent dimension fails to preserve important dynamic characteristics and inter-variable relationships, leading to reduced long-horizon prediction accuracy. In contrast, an excessively large latent dimension introduces unnecessary degrees of freedom, which can weaken generalization performance and reduce learning efficiency. The study further suggests that the optimal latent dimension depends on the complexity of the forecasting problem and tends to increase as the number of target variables grows. These findings highlight that latent-space capacity should be adaptively determined according to the dimensionality of the plant variables and the complexity of the prediction task.

This work contributes by emphasizing latent dimension optimization as a critical step in developing latent prediction models for nuclear power plant applications. The proposed perspective can support future applications such as trip margin prediction, mitigation action assessment, and advanced operator support for long-horizon, high-dimensional forecasting problems.