基于Adaptive Lasso与RF的航班运行风险预测改进研究

为了解决航班运行风险高维数组运算过于复杂的问题，同时为防止模型过度拟合影响预测精度，基于中国民航局发布的风险评估体系，以某航450组真实航班数据为标准样本，首先使用自适应套索算法（Adaptive Lasso）进行降维，从63项风险自变量中筛选出15项独立变量；然后，使用随机森林算法(Random Forest，RF)进行防过拟合处理，结果显示当使用重要度排序前12项变量拟合时，结果误差达到最小值，即得到最终预测指标；最后，构建Adaptive Lasso和RF的二阶段混合模型，同时选取主成分分析(Principal Component Analysis，PCA)、径向基函数(Radial Basis Function，RBF)网络、支持向量机(Support Vector Machine，SVM)3种对比方法，使用十折交叉验证精度.结果表明：Adaptive Lasso方法在筛选掉48项指标后，结果精度未见下降；经RF处理后4种方法评估精度均大于未处理前；Adaptive Lasso-RF混合模型的预测准确率和稳定性均优于PCA、RBF神经网络和SVM等方法.综上说明混合模型实现了有效降维和防过拟合，可大幅提升预测精度，用于解决航班风险预测问题可行并有效.

Flight Operations Risk Improvement Prediction Based on Adaptive Lasso and RF

In order to solve the problem of high computation complexity of flight operations high-dimensional arrays, and to avoid models over-fitting and the prediction precision influence, the 450 real flights data of certain airlines and the risk assessment system issued by the Civil Aviation Administration of China are used. Firstly, Adaptive Lasso is used to reduce dimensions, the fifteen independent variables are selected from 63 risk factors. Then, 12 items with the highest importance ranking and lowest error are retained according to the random forest over-fitting test results. Finally, the two methods are combined as a two-step hybrid algorithm; the principal component analysis, RBF neural network and support vector machine are selected for comparison and analysis; 10-fold cross validation is used to confirm outcome. The results show that the results accuracy does not change significantly after 48 indicators deleted by Adaptive Lasso method，and the assessment accuracy after RF processing is better than before RF. The prediction accuracy and stability of the Adaptive Lasso-RF hybrid model are better than that of principal component analysis, RBF neural network and support vector machine. All above conclusions indicate the Adaptive Lasso-RF prediction method is feasible and effective to solve the flight risk prediction problem, which can reduce dimensions, avoid models over-fitting and improve accuracy significantly.