Ensemble learning methodologies to improve core power distribution abnormal detectability

•Ensemble learning methods are proposed to improve performance of abnormal detectability.•Methods of PEM, TPS, WCM and ANN are used as base learners.•Boosting and bagging reduce errors for unstable learners; otherwise may increase them.•Stacking performs much better than bagging and boosting. Several ensemble learning methodologies have been presented to improve the performance of the core power distribution reconstruction, such as boosting technology, bagging technology and stacking technology. Four commonly used core power reconstruction methods, including PEM, TPS, WCM and ANN are adopted as base learners or meta-learners. The power distribution of the control rod drop accident and quadrant power tilt accident, which combine with detectors measurement noise, is reconstructed. The numerical simulation shows that the results of the stacking methodology can be better than any single one of these base learners under large sample spatial variation of power distribution and large measurement uncertainty. Otherwise results that close to the best one of these base learner results could be generated in stacking methods. In bagging and boosting methodologies, for the unstable learner such as TPS and ANN methods with large measurement uncertainty, both methodologies will help to reduce the reconstruction errors. The boosting methodology could even get better results than bagging. While for the stable learner (such as PEM learner, WCM learner) or unstable learner (e.g. the TPS or ANN method) but under small spatial variation (or small measurement uncertainty) that is not sensitive to the change of sample data, the bagging or boosting method even will reduce the reconstruction performance. Generally, ensemble learning methodologies have capability to improve the core abnormal detectability in most operation scene. Usually stacking methodology performs much better than bagging and boosting methodologies.