As a current focus of clean energy, nuclear energy has become an important part of the global energy structure transformation due to its high efficiency, low carbon and sustainability. As the main platform of nuclear energy utilization, the research focus on nuclear energy is the operation state monitoring and forecasting of the nuclear power plants (NPPs). NPPs contain a large number of multi-scale physical quantities and key components, and it is very important to study and forecast their operation states. With the full promotion of digitalization of nuclear facilities, the huge amount of operation data has provided the possibility of data-driven time series forecasting methods for NPPs. This paper develops a time series forecasting framework for NPPs, which is based on autocorrelation feature and Transformer architecture by using three models. First, dimension reduction model is built using Autoencoder architecture to select key variables for improving forecasting efficiency. Then, autocorrelation detection model based on LSTM architecture is used to extract autocorrelation features of the time series data. Finally, the time series forecasting model using Transformer architecture adjusted with autocorrelation is applied to forecast further time series data. The introduction of autocorrelation enhances the framework to learn internal dependence relationships within time series. This framework has been experimentally verified on a large-scale fault data set and has shown excellent performance in improving forecasting accuracy and model generalization ability.