Outage duration prediction under typhoon disaster with stacking ensemble learning

We propose a novel stacking ensemble learning model to predict the outage duration during typhoon disaster to help users prevent disasters. The model integrates extra tree(ET), extreme gradient boosting(XGBoost), light gradient boosting machine(LightGBM), random forest(RF), gradient boosting regression(GBR), decision tree(DT) as the base learner and GBR as the meta learner to enjoy the advantage of various accurate machine learning algorithms. First, the Batts wind field model is simulated to collect meteorological data. Geographical and power system data are also collected as the input sample. Then condensed nearest network(CNN) down-sampling and synthetic minority oversampling technique(SMOTE) algorithm over-sampling are used to preprocess the original data to solve the problem of unbalanced sample. Further, the Pearson correlation coefficient and model contribution are comprehensively analyzed to screen the final input characteristic variables. Next, the input characteristic variables are transmitted to the stacking ensemble learning model get trained to obtain comprehensive outage duration information. The scientificity and effectiveness are verified by a case study in Yangjiang City, Guangdong Province, China under No. 7 typhoon "Chapaka" in 2021. Simulation result shows the precision of the proposed stacking ensemble learning method is better comparing with any single algorithm (e.g., ET, RF, GBR).