AI-driven correction of near-surface wind speed forecasts in the Kumtag Desert: model comparison and key factor analysis

Accurate near-surface wind speed forecasting is critical for wind energy development, dust storm mitigation, and disaster risk management in arid desert regions. In the Kumtag Desert, forecasts from the ECMWF numerical weather prediction products often exhibit systematic biases due to complex local boundary-layer dynamics. This study investigates biases in near-surface wind speed forecasts from the ECMWF numerical weather prediction products in the Kumtag Desert. Using data from January 2020 to December 2022, three forecast lead-time groups (3–24 h, 27–48 h, 51–72 h) were constructed along with automatic weather station observations. Sensitivity experiments identified key input features, including 10 m height U/V wind components and 100 m vertical wind speed. Deep learning models (CNN2D/3D, DenseNet121, ResNet50/101 and KAN variants) and machine learning models (Gradient Boosting Regression, Elastic Net, Lasso, Ridge, Linear Regression, k-NN, Decision Trees, MLP) were evaluated using MSE, MAE, R², Pearson’s r, and classification metrics for winds above Beaufort scale level 5. Results show Gradient Boosting Regression consistently achieved the highest accuracy, improving R² and reducing biases in high wind speed forecasts. KAN integration was implemented by superimposing single-variable kernel approximation functions onto the output layers of ResNet/DenseNet. The selection of basis functions and hyperparameter settings was done empirically without systematic sensitivity analysis. This lack of explicit justification and optimization process limits reproducibility and makes the observed short-term improvement difficult to generalize. Factor importance ranking was U-10 m > V-10 m > V-100 m, with cumulative contribution above 80%, confirming the dominance of local boundary-layer processes. Contributions decreased with lead time but remained primary. The findings provide an effective approach for wind speed correction and gale warning in desert and arid regions, supporting applications in wind energy assessment, dust storm forecasting, and disaster risk reduction.