Bayesian-optimised hybrid machine learning model for coastal wind gust prediction in a marine-influenced atmospheric boundary layer

Abstract

Accurate prediction of wind gusts is crucial for applications in aviation, coastal and marine operations, and atmospheric dynamics research. This study presents a novel model combining a Sequencing Block and a Layer Perceptron (MLP) optimised using Bayesian Optimisation (B-MLP) to enhance the precision of coastal atmospheric wind gust forecasts. The model is validated using a 13-year dataset (January 2010 to March 2023) from Muscat International Airport, a coastal site influenced by Gulf of Oman sea–land breeze interactions. The Sequencing Block is designed and developed to capture the optimal arrangement of dataset segmentation using atmospheric and boundary layer parameters, thereby enhancing the model's predictive accuracy. The B-MLP model's efficacy is compared against traditional methods, including Decision Tree (DT) and Support Vector Regression (SVR), demonstrating a substantial enhancement in forecast quality. The B-MLP model achieves a correlation coefficient of 0.817 between actual and forecasted wind gusts, outperforming DT and SVR by notable margins in both accuracy and error reduction. The newly proposed model is validated using a 13-year dataset (January 2010 to March 2023) from Muscat International Airport, a coastal site influenced by Gulf of Oman sea–land breeze interactions, to prove its robustness and applicability on a 1-day ahead prediction horizon. The proposed B-MLP model improves forecast accuracy and offers a scalable solution for atmospheric boundary layer studies, marine safety applications, and real-time meteorological data analysis.