基于优化AC-BiLSTM模型的机场终端区流量预测

Airport Terminal Area Traffic Forecasting Based on Optimized AC-BiLSTM Model

将深度学习用于机场终端区流量预测研究。神经网络具有复杂的网络拓扑结构和较多的超参数,这使得超参数的选择很困难。以往神经网络模型的超参数通常是根据经验来调整的,因此,为了解决超参数搜索问题,引入粒子群算法(PSO)。针对基本粒子群算法(Basic PSO,BPSO)的缺点和不足,提出一种基于自适应惯性权重的混沌粒子群优化(AWCPSO)算法,该算法在基本PSO的基础上进行了改进和优化。具体方法是,使用一种新的动态自适应惯性权重在全局寻优和局部寻优之间取得平衡;进而将混沌思想和粒子群算法相结合,解决粒子群算法易陷入局部最优的问题。实验表明,AWCPSO算法用于AC-BiLSTM模型的超参数寻优时,既解决了BPSO算法的早熟收敛问题,又提高了流量预测的精度。

Deep learning is used for airport terminal area traffic prediction study.Neural networks have complex network topology and a large number of hyperparameters,which makes the selection of hyperparameters difficult.Previously,the hyperparameters of neural network models were usually adjusted empirically,so the particle swarm algorithm(PSO)was introduced to solve the hyperparameter search problem.To address the shortcomings and deficiencies of the basic particle swarm algorithm(Basic PSO,BPSO),a chaotic particle swarm optimization(AWCPSO)algorithm based on adaptive inertia weights is proposed.Specifically,a new dynamic adaptive inertia weight is used to strike a balance between global search and local search,and then the chaos method and particle swarm algorithm are combined to solve the problem that the particle swarm algorithm is prone to fall into local optimum.The experiments show that the AWCPSO algorithm solves the premature convergence problem of the BPSO algorithm and improves the accuracy of traffic prediction when used for hyperparameter search optimization of the AC-BiLSTM model.