In a large-scale plant such as a nuclear power plant, thousands of process values are measured for the purpose of monitoring the plant performance and the health of various systems. It is difficult for plant operators to constantly monitor all of the process values. We present a new data-driven method to monitor many process values and to enable early detection of anomaly signs including unknown events with few false detections. In order to accurately predict the process values in the normal state, we created a two-stage model composed of a time window autoencoder and a deviation autoencoder. The two-stage model handles a large number of process values, their rapid changes of the process values such as an operation mode change, changes of the process values in both the steady and the transient states, and the external disturbances such as exogenous noise, atmospheric temperature, etc. The time window autoencoder examines time correlations of time series process values while the deviation autoencoder treats correlations of variation due to external factors. We evaluated a predicting ability of the rapid changes, detection performances in the transient state, and detection performances under noisy conditions with simulated process values of a nuclear power plant, a 1,100 MW Boiling Water Reactor having 3,100 analog process values. The two-stage model clearly showed a good anomaly detection performance with zero or few false detections. The two-stage model would be an effective solution for plant monitoring and early detection of anomaly signs.