Forecast of severe typhoon formation based on unbalanced data

Abstract

As extreme weather events, severe typhoons have a great impact on people's daily production and social stability. Therefore, many scholars have used different methods to predict the formation of severe typhoons in order to mitigate the impact of severe typhoons. However, the severe typhoon in reality is a small probability event and the number of severe typhoon samples used for prediction is much smaller than the number of ordinary typhoon samples. Existing models rarely consider the impact of unbalanced data on model training and prediction, which makes it difficult to apply the model to the actual severe typhoon prediction. Therefore, we propose a severe typhoon formation prediction model based on unbalanced data. The model uses a convolutional neural network to obtain features from the severe typhoon environmental field and an LSTM model to implement the prediction of severe typhoon formation. A customized loss function designed in this paper is used to add different classification weights for normal typhoon samples and severe typhoon samples, so that the model improves the prediction of unbalanced severe typhoon formation. The experiments show that the severe typhoon formation prediction model based on unbalanced data outperforms the traditional machine learning model for unbalanced severe typhoon formation prediction.