Development of Prediction Model for Damage Costs of Heavy Rainfall Disasters Using Machine Learning in the Republic of Korea

In this study, a prediction model was developed that considers the rainfall characteristics and damage characteristics of heavy rainfall disasters in Korea using machine learning models. Considering the damage characteristics of heavy rainfall disasters that occurred from 1999 to 2019 in 228 administrative districts in Korea, four types of total rainfall and five types of damage costs were selected to predict the total damage cost. The machine learning models selected for this study were Random Forest, K-Nearest Neighbors, Decision Tree, and eXtreme Gradient Boosting, and their accuracy was evaluated using R2, EVS, and MAPE. The training period spanned from 1999 to 2015, while the evaluation period extended from 2016 to 2019. The Random Forest model emerged as the most effective model for predicting the total damage costs associated with heavy rainfall disasters, exhibiting an accuracy of 0.95 for R2, 0.95 for EVS, and 0.05 for MAPE. It was observed that when the total damage costs are minimal, all models demonstrate high prediction capability. However, as the damage costs escalate, the prediction power experiences a decline due to the presence of errors. The machine learning prediction model for heavy rainfall disasters developed in this study has the potential to contribute to national efforts aimed at preventing and preparing for heavy rainfall disasters.