Classification prediction of load losses in power stations using machine learning multilayer stack ensemble

Load losses negatively impact the reliability of power stations, leading to plant failures. To support the decision-making of improving plant reliability, we experimented with six machine learning classifiers to find the model combination that produces the best prediction performance, called the Explainable Multilayer Stack Ensemble. We applied a five-year dataset from six power stations. Since the dataset is highly imbalanced with the positive class dominant, class weights are calculated and assigned to reduce bias toward the majority class. The best parameters are determined through a randomized search with cross-validation and applied to train the models. The Explainable Multilayer Stack Ensemble performed better than the individual models, with a further improvement by excluding the Gaussian Naïve Bayes in the second layer since it produced high false negatives. We demonstrate that when handling a highly imbalanced dataset, balanced accuracy, Receiver Operating Characteristics, and Precision-Recall Area Under the Curve provide a more reliable evaluation of model performance than focusing solely on standard evaluation metrics, such as accuracy, precision, and recall. Moreover, by excluding a poor-performing classifier from ensemble, we optimized the prediction process, and further enhanced overall performance.