【摘要】 实现城市对外客运枢纽抵站客流的精准预测，是增强枢纽接续运输运力调度科学性的重要前提。为提高枢纽抵站客流的预测精度，研究了基于超参数优化的鲸鱼算法与双向长短期记忆神经网络模型（whale optimization algorithm and bi-directional long short-term memory,WOA-Bi-LSTM）组合的客流预测方法。融合历史抵站客流数据及天气、日期、时段等多源信息，分析抵站客流的时变特性，并开展不同影响因素与枢纽抵站客流量间的相关性分析。改进了传统双向长短期记忆神经网络模型（bi-directional long short-term memory,Bi-LSTM）的参数设置方法，用鲸鱼算法（whale optimization algorithm,WOA）代替手动调参，选取学习效率（η）与隐藏神经元个数（H）2个对模型预测精度具有较大影响的超参数进行最优超参数组合搜寻，通过计算其适应度函数进行循环逻辑判断，实现参数自适应优化。通过不断寻优，获取最优参数组合值，确定设置η为0.060 3、H为120，并输出预测结果和3个模型精度评价指标（R~2判定系数，平均绝对误差与均方根误差）；同时构建了3种不同超参数优化算法改进的Bi-LSTM组合模型、2种基于WOA算法改进的其他组合模型，以及2种未改进的神经网络模型与WOA-Bi-LSTM模型使用相同的抵站客流数据集进行多维度对比，验证所建模型的优越性与鲁棒性。结果表明：WOA-Bi-LSTM模型在节假日、工作日与非工作日等不同枢纽抵站客流预测场景下均体现出良好的适用性，与其他模型相比，R~2相关系数最大，达到0.951 4，表示所建模型的拟合效果最好；平均绝对误差与均方根误差最小，分别为762.96与556.25，误差相较于其他模型至少减少5.6%和3.2%。更多还原

【Abstract】 Accurate prediction of arrival passenger flows at external passenger transportation hubs is an important prerequisite for enhancing the scientific scheduling of the transferring transport capacity of hubs. In order to improve the prediction accuracy of arrival passenger flows, a combination model of the whale optimization algorithm and bi-directional long short-term memory(WOA-Bi-LSTM) is proposed. Integration of historical arrival passenger flow data with multi-source information such as weather, date, and time of day, the time-varying characteristics of arrival passenger flows are analyzed, and correlation analysis is conducted between different influencing factors and arrival passenger flows at the hub. The parameter setting of the traditional bi-directional long short-term memory(Bi-LSTM) model is modified with the whale optimization algorithm(WOA) optimization algorithm. Learning rate(η) and the number of hidden neurons(H) are significant hyperparameters on model prediction accuracy and are determined by searching optimal values. The search procedure is performed to achieve adaptive parameter optimization by calculating their fitness functions through iterative logic. Through continuous optimization, set the η as 0.060 3 and H as 120. The performance of the proposed model is evaluated using three indicators: R~2 value, mean absolute error(MAE), and root mean square error(RMSE). Simultaneously, the WOA-Bi-LSTM model is compared with several baseline models across multiple dimensions based on the same dataset, including three Bi-LSTM models modified by different hyperparameter optimization algorithms, two other combination models based on the WOA algorithm and two unmodified neural network models. The results show that the WOA-Bi-LSTM model shows better performance of predicting arrival passenger flows in different scenarios involving holiday, workday and non-workday. Compared to other models, the WOA-Bi-LSTM model achieves the highest R~2 of 0.951 4, indicating that the proposed model has the best fit. The RMSE and MAE are both the lowest, at 762.96 and 556.25, respectively, with errors reduced by at least 5.6% and 3.2% compared to other models.更多还原