Enhancing Grid Reliability: Predicting Power System Faults Using the Multivariate Generalized Dirichlet-Multinomial (GDM) Model for Correlated Count Data

This research has developed a multivariate regression model to predict the occurrence of faults and identify the factors influencing fault occurrence in three feeders of the Provincial Electricity Authority, Rangsit Branch, Thailand, in 2024. The most appropriate multivariate distribution function has been selected in cases where the dependent variables have been related. Multivariate regression models have been created for count data consisting of the Dirichlet-Multinomial (DM) model, the Negative Multinomial (NegMN) model, and the Generalized Dirichlet-Multinomial (GDM) model. The GDM model has best fitted a real data set with an overdispersion problem and a large number of zero values (AIC value = 103.914, BIC value = 107.77), followed by the DM model (AIC value = 104.157, BIC value = 108.554) and the NegMN model (AIC value = 277.474, BIC value = 285.277). The most efficient regression model has been the GDM model (AIC value = 111.195, BIC value = 131.715), followed by the DM model (AIC value = 238.836, BIC value = 281.7633) and the NegMN model (AIC value = 238.836, BIC value = 281.7633). The factors that have significantly affected the occurrence of faults in power systems at the 0.05 statistical level have been the number of days with strong winds each week (X2), the number of days with humid weather each week (X3), the number of fallen trees/natural disasters detected each week (X4), and the number of times that equipment failures have been detected each week (X6).