Predicting Algorithm of Thunderstorm Days in the Northern Region of Chile Using Convolution Neural Network

Abstract

Advances in predicting thunderstorms have been made possible through the use of artificial intelligence. Convolution neural networks, inspired by the processes of the human brain, are particularly effective in image classification. In particular, one-dimensional convolution neural networks have played a significant role in time series analysis, including thunderstorm forecasting. Unfortunately, these models face challenges when used with unbalanced datasets, where the proportion of events of interest, such as thunderstorms, is significantly lower than other phenomena. To overcome this limitation, several over-sampling and under-sampling strategies have emerged. In this paper, we propose a method for forecasting thunderstorm occurrences in the northern region of Chile using a one-dimensional convolutional neural network, combined with a balanced batch generator and attention models. The algorithm developed to predict thunderstorm days achieved a performance metric of approximately 79.6%, a promising result due to the minimal failure rates observed.

Keywords

• Artificial neural network
• machine learning algorithms
• thunderstorm days