Research on CANDU Channel Power Prediction Based on Stacking Ensemble Learning

The accuracy of CANDU reactor channel power prediction directly affects the quality of the refueling plan, which in turn affects the economy and safety of reactor operation. To improve the prediction effect, it is advocated to introduce artificial intelligence algorithms to mine the potential variable relationship from historical operation data. After data cleaning and feature selection, a feature termed "Refueling Impact Index" is designed. With XGBoost, random forest, support vector regression, and BP neural network as primary learners and support vector regression as secondary learners, an ensemble learning model based on Stacking is constructed. Through comparative analysis, the Stacking ensemble learning model has achieved a "secondary improvement" in prediction effect on the basis of a single learning model. Moreover, the Stacking ensemble learning model has significantly better effect than the traditional RFSP methods in terms of average prediction deviation rate, maximum prediction deviation rate, and prediction deviation rate variance. This enables physical engineers to obtain more accurate power feedback in the process of formulating refueling plans, scientifically select refueling channels, and thereby improve economic benefits while ensuring reactor safety.