In a large-scale plant such as a Nuclear Power Plant (NPP), thousands of process values are measured for the purpose of monitoring the plant performance and the system health. It is difficult for plant operators to constantly monitor all of the process values. We present a data-driven method to comprehensively monitor a large number of process values and detect early signs of anomalies, including unknown events, with few false positives. In order to learn the complex changing internal state of a NPP and accurately predict the normal process values, we have developed a two-stage autoencoder (TSAE), a type of neural network, composed of a time window autoencoder and a deviation autoencoder. TSAE realizes to detect anomalous signals during the plant transient conditions by collecting time-series data and learning the nonlinear temporal correlation among them. In the actual plant, some process values which are physically uncorrelated with each other happen to behave similarly (pseudo-correlation). Learning the pseudo-correlation by the algorithm causes false positives because the predicted values of unrelated process values are incorrectly correlated. Therefore, Toshiba has proposed the model classification concept of separating the process values into two groups based on physical correlation and applied a model structure of TSAE. As a result, it becomes possible to learn only with the process values that are physically correlated and enhance the performance of prediction/detection. We assessed the improved TSAE with simulated process values of a NPP and showed excellent performances with few false positives.