基于多任务NR-DenseNet网络的航班延误预测模型

Flight Delay Prediction Model Based on NR-DenseNet

不同于目前大多数只倾向于研究单一的分类或回归任务的航班延误预测方法,该文提出一种基于多任务NR-DenseNet网络的航班延误预测模型,旨在同时实现航班延误等级分类预测与延误时间回归预测。首先,预处理相关数据;其次,建立多任务学习特征提取共享层,使用NR-DenseNet网络提取任务之间的共享参数,深度挖掘任务之间的相关特征;然后,建立多任务学习特定任务层,通过回归器与分类器分别输出特定任务的预测结果;最后,采用损失加权方法对两个任务损失函数进行优化,平衡任务间的收敛速度,提高模型泛化性。将模型应用在宁波机场数据集中,与单任务模型相比回归任务平均MSE降低了23.4%,平均MAE降低了14.2%,分类平均准确率提升了2.7%。实验结果表明,该文方法提升了分类任务的准确率降低了回归任务的误差,可以有效提升模型性能。

Different from most of the current flight delay prediction methods that only tend to study a single classification or regression task,a flight delay prediction model based on multi-task NR-DenseNet was proposed in this paper,which aims to achieve both flight delay level classification prediction and delay time regression prediction.Firstly,the relevant data was pre-processed;Secondly,multi-task learning feature extraction sharing layer was established,NR-DenseNet was used to extract shared parameters between tasks to dig deeper into relevant features between tasks;Then,a multi-task learning task-specific layer was established to output the prediction results of specific tasks through a regressor and a classifier respectively;Finally,the loss-weighting method was used to optimize the loss functions of the two tasks,balance the convergence rate between tasks,and improve the generalization of the model.Compared with the single-task model,the average MSE of the regression task was reduced by 23.4%,the average MAE was reduced by 14.2%,and the average classification accuracy was increased by 2.7%.Experimental results show that the proposed method improves the accuracy of the classification task and reduces the error of the regression task,which can effectively improve the model performance.