Wind speed plays a central role in the assessment of wind power resources. This study begins with spatial analysis using Inverse Distance Weighting (IDW) and kriging techniques applied to truncated Weibull-distributed wind speed data. To capture both spatial and temporal dependencies, a spatio-temporal modelling approach was implemented using the Gradient Boosting Method (GBM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). Among these, XGBoost achieved comparatively higher R² and lower Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) and was selected for further optimization. However, the R² values remained low across all models, indicating limited predictive power. To address this, model performance was enhanced through data preprocessing, feature engineering, and parameter tuning. Because wind speed data often display skewness, this study designed a tailored Weibull deviance loss function and incorporated it into the XGBoost algorithm. A simulation experiments was carried out to evaluate its performance against the commonly used default loss functions. The findings showed that the customized loss consistently produced lower Mean Absolute Error (MAE) and Gamma deviance values. This outcome indicates that adopting a Weibull deviance loss is well-suited for data following a Weibull distribution and can serve as an effective approach to improving the predictive accuracy of wind speed models built with XGBoost. Specifically, this study aims to characterize the spatial distribution of truncated Weibull wind speeds using IDW and kriging techniques, and to examine whether a Weibull-based deviance loss improves the predictive accuracy of XGBoost relative to its default regression objectives under skewed wind speed distributions. • Truncated Weibull wind speed data were spatially analyzed using IDW, kriging, and block kriging across Bangladesh. • Spatio-temporal wind speed prediction was performed using RF, GBM, and XGBoost machine-learning models. • XGBoost performance was improved through feature engineering, parameter tuning, and temporal–spatial encoding. • A customized Weibull deviance loss function was developed and integrated into the XGBoost framework. • Simulation studies confirmed the superiority of Weibull deviance over standard XGBoost regression objectives.
Jahan et al. (Mon,) studied this question.