Mental workload (MWL) assessment is crucial for mitigating human errors during long-term complex operation tasks in nuclear power plant (NPPs). EEG is recognized as one of real-time, objective and effective MWL assessment methods. This study aimed to investigate the roles of brain regions and temporal variation of EEG signals in EEG-based MWL assessment during long-term complex NPP operation tasks. EEG signals were collected from sixteen operators who performed simulated long-term complex operation tasks. Four typical machine learning algorithms were employed to develop MWL assessment models. The results showed that the temporal region achieved the best accuracy (88.36%) across four brain regions, while whole-brain regions reached an accuracy of 91.07%. Model performance mostly improved when temporal variation of signals was considered. This study demonstrated the effectiveness of temporal region and the beneficial role of temporal variation of EEG signals in operators’ MWL assessment. The findings provide important implications for developing cost-effective MWL assessment models during long-term complex operation tasks in NPPs.