Research on Flood Prediction and Early Warning System Based on Machine Learning Models

This document outlines a flood prediction system that utilises multiple machine learning models. Through Lasso regression and its combinations with linear regression, as well as with Gradient Boosting Regression Tree, the main predictors of flood occurrence such as deforestation, population scores, and coastal vulnerability were captured. An early warning system able to classify flood risks into three levels (high, medium, and low) was created utilising K-means clustering combined with Bayesian optimisation. The model achieved best results in the Bayesian-optimised Support Vector Regression (SVR) model with MSE of 0.00036269 and R-squared of 0.86073. We further enhanced the model by applying Principal Component Analysis (PCA) for feature reduction followed by hyper-parameter tuning with the Dung Beetle Optimiser (DBO) algorithm. The final distribution analyses of prediction results revealed non-normal features using the Jarque-Bera test. The entire system is designed to improve the accuracy of flood predictions and provide scientific evidence for better planning of flood prevention and disaster reduction efforts.