Accurate recognition of operators’ mental workload is essential for ensuring system safety and operational efficiency in long-term and complex nuclear power plant tasks. However, developing lightweight and effective mental workload recognition models remains a major challenge. This study aims to construct an effective EEG-based workload recognition model by exploring the roles of feature combinations from different brain regions. EEG data were collected from 10 nuclear power plant operators. Four typical machine learning algorithms were used to build classification models. The results show that different brain regions contributed variably to mental workload recognition, with the temporal lobe yielding the best performance across single brain regions. Using EEG features from only the temporal, parietal, and occipital lobes achieved classification accuracy comparable to that of full-brain features, while significantly reducing feature dimensionality and system complexity. These findings support the development of lightweight and practical EEG-based mental workload monitoring systems.