低空湍流具有显著局地性与时空间歇性，传统预测模型难以精准预测，因此本文提出了一种混合专家模型，实现低空湍流预测。为捕捉湍流样本的全局规律，该混合专家模型中的基线专家基于时间卷积网络学习全量样本的动态变化趋势；为实现极端湍流样本非线性特征的差异化建模，残差专家通过参数敏感性分析实验筛选极端湍流样本，且以基线预测值与真实值之间的残差作为学习目标，修正预测偏差。同时，为缓解固定权重的双专家融合策略难以适配不同湍流模式的问题，设计以多特征融合为输入的学习门控，通过多层感知机结构生成动态权重对双专家预测结果进行自适应融合，从而实现全量样本预测结果稳定性与强湍流高值样本预测精度。本研究以新疆喀什机场和兰州中川国际机场的激光雷达观测数据开展实验，该方法整体准确率分别为96.5%和92.7%。所提方法为低空湍流的实时预测与航空安全预警提供可参考的思路。

Low-altitude turbulence exhibits significant locality and spatiotemporal intermittency, making accurate prediction challenging for traditional forecasting models. Therefore, a hybrid expert model is proposed in this study to achieve low-altitude turbulence prediction. To capture the global patterns of turbulence samples, the baseline expert in the hybrid expert model leverages a temporal convolutional network (TCN) to learn the dynamic variation trends of the entire dataset. For the differentiated modeling of nonlinear characteristics of extreme turbulence samples, the residual expert screens extreme turbulence samples through parameter sensitivity analysis experiments, and takes the residual between the baseline predicted values and the true values as the learning target to correct prediction deviations. Meanwhile, to address the issue that the dual-expert fusion strategy with fixed weights struggles to adapt to different turbulence patterns, a learning gate with multi-feature fusion as input is designed. Dynamic weights are generated through a multilayer perceptron (MLP) structure to adaptively fuse the prediction results of the two experts, thereby achieving both the stability of prediction results for all samples and the prediction accuracy for high-value severe turbulence samples. Experiments are conducted using lidar observation data from Kashgar Airport in Xinjiang and Lanzhou Zhongchuan International Airport, and the overall accuracy of the proposed method reaches 96.5% and 92.7%, respectively. The proposed method provides a referable approach for the real-time prediction of low-altitude turbulence and aviation safety early warning.