Abstract
Floods in 2022 caused significant economic losses in Malaysia, totaling RM 6.1 billion ($1.46 billion). Selangor, a densely populated and industrialized state, was the hardest hit, with the manufacturing sector suffering RM 900 million in losses. Accurate rainfall prediction is essential for effective weather forecasting and climate modeling in the region. This study evaluates the effectiveness of a hybrid Statistical Downscaling-based Hidden Markov Model-Machine Learning Model (SD-based HMM-ML) for rainfall prediction. It also explores optimal imputation methods for handling missing data, selects predictors using dimensionality reduction, and addresses uncertainties in zerobounded rainfall data. Local rainfall (predictand) and atmospheric data (predictor) from 33 stations in Selangor (2008_2018) were analyzed. Seven imputation methods were tested: Mean, Median, Expectation-Maximization (EM), Markov Chain Monte Carlo (MCMC), k-Nearest Neighbor (kNN), Non-iterative Partial Least Square (NIPALS), and Random Forest (RF). Principal Component Analysis (PCA) reduced highdimensional data, selecting five principal components with a high cumulative variance. HMM addressed zero-bounded rainfall uncertainties, identifying three hidden states with the lowest Bayesian Information Criterion (BIC). Five hybrid models _ HMM-RF, HMM-SVM, HMM-DT, HMM-KNN, and HMM-ANN _ were evaluated using RMSE, MAE, MFE, and NSE. Median Imputation had the lowest values of RMSE and MAE, and highest value of NSE across all stations. HMM-RF outperformed other models, demonstrating superior accuracy. By leveraging machine learning, this novel hybrid model enhances rainfall prediction accuracy, improving early warning systems, infrastructure planning, and flood mitigation efforts. The study contributes to building a more resilient urban environment, mitigating economic losses from future floods in Selangor.