Empirical-singular-wavelet based machine learning models for sea level forecasting in the bosphorus strait: A performance analysis

Abstract

Forecasting sea level is significant for sustainable water supply management, flood mitigation, shoreline maintenance, ecological sustainability, and economic advancement. In this study, a novel approach that incorporates decomposition methods with machine learning algorithms is presented. The objective is to enhance the accuracy of sea level predictions by tackling the challenges associated with sea level modelling, which arise from its complex characteristics and influence of various factors. This study introduced predictive frameworks incorporating Wavelet Decomposition (WD), Singular Spectrum Analysis (SSA), Empirical Mode Decomposition (EMD) with XGBoost (XGB), Decision Trees (DT), k-Nearest Neighbors (KNN), and Support Vector Machines (SVM). The study conducted predictions for time spans of 1, 7, and 30 days, evaluating the accuracy of the model's predictions using performance metrics. For the first day prediction, the XGB model achieved highest NSE value of 0.86, whereas the hybrid model (SSA-XGB) attained NSE value of 0.99. The hybrid model (EMD-XGB) produced remarkable results, attaining NSE value of 0.70 for the 7 day prediction time horizon. Additionally, the hybrid models (EMD-XGB and EMD-KNN) remarkably achieved NSE values exceeding 0.60 for 30-day time horizon. The results demonstrate that the hybrid models accurately predict daily sea levels with a lead time of up to 7 days. Additionally, the hybrid models demonstrated satisfactory prediction performance for time spans of up to 30 days. The hybrid models presented enhancements in both prediction accuracy and extended lead time predictions.