Abstract: This paper proposes a hybrid optimized model named SSLT-GRU, which integrates Sparrow Search Algorithm (SSA) with Gated Recurrent Units (GRU) into an LSTM-Transformer architecture. The model first employs LSTM to capture long-term temporal dependencies in the wind speed sequences. Subsequently, a Transformer module is introduced to model global dynamic correlations, thereby enhancing the representation of complex disturbances such as wind shear. Moreover, the SSA algorithm is utilized to cooperatively optimize key hyperparameters of the LSTM-Transformer component, mitigating the risk of manual tuning being trapped in local optima. Additionally, a residual GRU module is incorporated to refine the initial forecasts by correcting residual errors, further improving forecasting accuracy. Experimental results on two real-world wind speed time-series datasets collected from different aerial platforms demonstrate that the proposed model achieves a root mean square error (RMSE) of 0.296 m/s and 0.4801 m/s, and a mean absolute percentage error (MAPE) of 5.69% and 3.58% for 6-step-ahead forecasts, respectively. For 12-step-ahead predictions, the RMSE values are 0.3539 m/s and 0.6856 m/s, with MAPEs of 6.89% and 6.45%, respectively. These results validate the superior robustness and forecasting accuracy of the proposed model for low-altitude wind speed forecasting.