Prediction of thunderstorm days in Chilean territory using machine learning techniques

Abstract

This study presents the design of a model that allows for the prediction of thunderstorm days. The study was conducted in a location within Chile, with the requirement that it be an area that has recorded a high number of lightning and has sufficient meteorological records between the years 2012 and 2021 to train a machine learning model. The chosen location was Visviri, in the Arica and Parinacota region. A methodology was employed, which included creating a dataset, conducting Exploratory Data Analysis, handling missing data, feature engineering, feature selection, hyperparameter tuning, and sensitivity analysis to find the best-performing model based on the F1 score. The model developed was a multilayer neural network with ReLU activation function and dropout, achieving a performance of 74% in F1 score for the year 2021.