Monte Carlo simulation of lightning strike incidence and subsequent flashovers on medium voltage distribution lines, with or without the presence of shield wire(s). Flashover(s) due to five different modes of lightning interaction with the overhead line are analyzed: (1) direct strike to phase conductor w/o shield wire presence, (2) direct strike to phase conductor w/ shield wire present (i.e. shielding failure), (3) direct strike to shield wire with subsequent flashover (i.e. backflashover event), (4) indirect (i.e. near-by) lightning strike w/ shield wire(s), and (5) indirect lightning strike w/o shield wire(s). Indirect strikes provoke flashovers on the transmission line insulation by means of the induced overvoltage (from electromagnetic coupling). Three different models of induced voltage computation, from the indirect lightning strikes, have been implemented: (a) Rusk's method, (b) Chowdhuri-Gross method, and (c) Liew-Mar method. Also, two different methods of backflashover analysis have been implemented. Generated data set is hierarchical in nature, and class imbalanced.
Machine learning application to classification of flashover incidents on distribution lines emanating from lightning strikes. Dataset imbalance is dealt with by using a stratified shuffle split for crating training and test sets, and by class weighting during model training. Support vector machine (SVM) classifier has been implemented as a base estimator for the bagging ensemble (i.e. the bootstrap aggregating ensemble of SVM base models). The bagging ensemble has been crafted using the off-the-shelf model (Variant A), as well as with a hand-made model (Variant B). Main difference between these variants emanates from the treatment of the base estimators. Hyperparameters optimization in each of the variants can be tackled by means of: (a) randomized search, (b) halving random search, and (c) grid search algorithms, in combination with a stratified k-fold cross-validation on the training data set. Pipeline is used for stringing together features processing with bagging ensemble creation and hyperparameters optimization. Models have been developed using the scikit-learn library. Jupyter Notebook can be seen rendered here.