飞机颠簸是影响飞机安全以及旅客舒适度的一种重要航空危险天气事件之一，因此，对飞机颠簸的精准预测尤为重要。虽然近年来已有采用数值天气预报（numerical weather prediction, NWP）和人工智能方法进行飞机颠簸研究和业务应用，但这些方法并未考虑颠簸随时间和空间变化的动态性，难以有效应对小尺度颠簸的复杂多变性。为此，使用机载设备观测的湍能耗散率（energy dissipation rate, EDR）数据，构建了一种CNN-Transformer-BiLSTM（CNTFBL）的深度学习模型，对飞机颠簸的预测进行了研究。对比试验和消融实验的结果表明，相比于传统的机器学习方法，该模型综合了CNN提取小尺度颠簸局部空间特征、Transformer捕捉长时间步长序列全局依赖关系、BiLSTM处理短时间的局部依赖关系的优势，在小尺度和瞬时性的中重度飞机颠簸预测中，准确率得到了较大的提高，尤其在预测时间步长较长时效果比较明显。当提前9 h和12 h对飞机颠簸进行预测时，其准确率分别高达0.950 3和0.953 5。

Aircraft turbulence is one of the critical aviation weather hazards affecting flight safety and passenger comfort. Therefore, accurate predictions of aircraft turbulence are particularly essential. Although numerical weather prediction (NWP) and artificial intelligence methods have been applied in recent years for turbulence research and commercial applications, these approaches have not fully addressed the dynamic nature of turbulence across temporal and spatial changes, making it challenging to handle small-scale turbulence that is complex and variable. To address this, energy dissipation rate (EDR) data was used to develop a deep learning model, CNN-Transformer-BiLSTM (CNTFBL), aimed at turbulence prediction for aircraft. Comparison and ablation experiments show that, compared to traditional machine learning methods, this model leverages CNN to capture small-scale, localized spatial turbulence features, Transformer to capture long-term global dependencies, and BiLSTM to manage short-term local dependencies. These combined advantages significantly enhance accuracy in small-scale and short-term moderate turbulence predictions. Notably, the model performs exceptionally well at critical forecast times, achieving accuracy rates of 0.950 3 and 0.953 5 for turbulence predictions at 9 h and 12 h in advance, respectively.