Hybrid double ensemble empirical mode decomposition and K-Nearest Neighbors model with improved particle swarm optimization for water level forecasting

Abstract

Water level forecasting plays a vital role in environmental protection and flood management because reliable predictions allow for the deployment of early warning systems to alert the public to minimize the impacts of flooding. This study presents an enhanced approach for weekly water level and flood prediction by integrating data decomposition techniques with machine learning models. Specifically, Ensemble Empirical Mode Decomposition (EEMD) was applied to disaggregate the original water level data into distinct Intrinsic Mode Functions (IMFs) to simplify complexity and enhance periodicity detection. A secondary decomposition was performed on the high-frequency IMF 1, derived from EEMD, to further refine the data features. The K-Nearest Neighbor (KNN) and Support Vector Machine (SVM) models, optimized using Improved Particle Swarm Optimization (PSO), were employed for forecasting. The effectiveness of these hybrid models was evaluated using various performance metrics, revealing that the DEEMD-KNN-PSO and DEEMD-SVM-PSO models significantly outperformed other single decomposition and standalone models. Among these, the DEEMD-KNN-PSO model demonstrated superior accuracy in predicting water levels, showcasing its potential for reliable flood prediction in the Klang River region of Sri Muda, Malaysia. This approach highlights the value of data decomposition and machine learning optimization for improving water level prediction accuracy.

Keywords

• Artificial Intelligence
• Machine Learning
• Flood Forecasting
• Predictive Analytics
• Water Level